The withOAuthRetry function retries requests on 401 errors but has no maximum retry count or exponential backoff ceiling. This could cause repeated OAuth token refresh attempts against paid API endpoints, potentially draining budget through failed request charges.

The createCodeSession function accepts a baseUrl parameter that is directly concatenated into the request URL without validation. An attacker could inject malicious URLs or protocol handlers (e.g., javascript:, data:) through this parameter.

The fetchRemoteCredentials function accepts a baseUrl parameter that is directly concatenated into the request URL without validation. Same injection risk as createCodeSession.

EnvLessBridgeConfig handles bridge sessions with retry logic (init_retry_max_attempts: 3) but no per-tenant or per-user spend caps. Bridge operations may trigger paid API calls (session creation, archive uploads) with no budget enforcement, allowing unlimited cost accumulation.

resolveInboundAttachments() fetches files from bridge API without file size limits, download rate limiting, or per-user quotas. An attacker could trigger massive file downloads, consuming bandwidth and storage costs.

The bridge initialization checks OAuth token expiration but persists dead token state with failCount. If multiple processes fail to refresh tokens simultaneously, they may all attempt to use the same expired token, potentially causing credential leakage through repeated 401 responses to the auth server.

Session access tokens are logged in debug output with 'accessToken=present' or 'accessToken=MISSING'. This exposes credential information in logs without proper masking or consent.

Debug logs include sessionId, sdkUrl, and accessToken in a single log line. Session identifiers are PII that should not be logged without consent and should be masked.

The BridgeConfig type accepts apiBaseUrl and sessionIngressUrl as user-provided strings without validation. These URLs are used for HTTP polling and WebSocket connections respectively. Attackers could supply malicious URLs to perform SSRF, redirect traffic to internal services, or exfiltrate data.

The mcp list command outputs all configured MCP server URLs and connection types to console. This reveals internal service infrastructure and network topology.

The mcp get command outputs server URLs, OAuth client IDs, callback ports, and environment variables to console. This reveals internal service endpoints and authentication configuration to anyone with access.

mcpAddJsonHandler accepts user-provided JSON configuration for MCP servers without verifying the source integrity or validating against a trusted registry. User can inject arbitrary MCP server configurations that could compromise the system.

**Perspective 1:** jsonParse(line) is called without depth or size limits, allowing entity expansion attacks (Billion Laughs) that could cause memory exhaustion or denial of service through exponential XML/JSON expansion. **Perspective 2:** The line parameter passed to jsonParse() has no maximum length validation, allowing extremely long input strings that could exhaust memory during parsing.

The `alwaysValidStatus()` callback accepts ALL HTTP status codes as valid (returns true unconditionally). This bypasses axios's default error handling, potentially allowing the client to process error responses (4xx, 5xx) as successful operations. This could mask server-side failures and allow attackers to probe for vulnerabilities.

The getTransportForUrl function accepts a URL parameter that is used directly to construct SSE or WebSocket connections. If the URL comes from user input, it could be manipulated to connect to malicious endpoints.

The bridge command enables remote-control sessions connecting this terminal. The isEnabled() function only checks feature flag and isBridgeEnabled() - no authentication, authorization, or session validation is visible in this entry point.

addMarketplaceSource() accepts marketplace sources (GitHub, HTTPS, local paths) without cryptographic signature verification. Malicious actors could publish tampered marketplace manifests that appear legitimate.

GitHub URL is constructed from githubRepo_0 which comes from plugin metadata. If plugin metadata is user-controllable, this could enable SSRF through crafted repository names.

The importGithubToken() function passes the GitHub token to an external API endpoint. While the RedactedGithubToken class masks the token in string representations, the actual token value is sent in the request body. There's no verification that the receiving endpoint is legitimate or that the token won't be logged or exposed in error responses.

Plugin is loaded and executed without any cryptographic signature verification or provenance tracking. The plugin's install path is used directly without verifying the artifact's integrity or origin.

The upgrade command loads upgrade functionality without verifying the integrity of downloaded artifacts. Upgrade packages could be tampered with to inject malicious code or backdoors.

AutoUpdater fetches latest version from external API without verifying the source integrity or validating the update against a known-good signature. An attacker who compromises the update server could inject malicious code.

Update process lacks provenance tracking. There's no record of which update was applied, from which source, or when. This makes it impossible to audit or trace malicious updates.

Settings configuration is persisted via saveGlobalConfig() without encryption. Violates SOC 2 CC6.1 (data protection) and PCI-DSS 3.4 (protection of stored data). User settings may contain sensitive configuration including API keys, authentication tokens, and workspace paths.

**Perspective 1:** Session ID is logged in analytics event 'tengu_status_line_mount' without explicit consent tracking or user opt-out mechanism. Session identifiers are PII that should not be sent to analytics without consent. **Perspective 2:** Status line configuration settings including command and padding are logged to analytics. These settings may contain user preferences that constitute PII and should not be tracked without explicit consent.

The performMCPOAuthFlow is called with transport type cast from agentServer.transport without validation. This could allow HTTP transport to be used for OAuth flows, potentially exposing credentials over unencrypted connections.

The authorizationUrl is set from performMCPOAuthFlow without validation that it uses HTTPS. This could allow attackers to intercept OAuth tokens via man-in-the-middle attacks.

The authorization URL is constructed without validating the scheme. An attacker could potentially inject a malicious URL scheme (e.g., javascript:, data:) if the URL construction is not properly sanitized, leading to credential interception or code execution.

The OAuth flow constructs URLs without validating the transport type. If the base URL is not properly validated, it could allow redirect to malicious endpoints or protocol downgrade attacks.

The file_path from tool input is used directly in readFileSync and FileWriteToolDiff without checking for path traversal attempts. An attacker could use ../ sequences to access files outside the intended directory.

The filePath parameter from sedInfo is used directly with getFsImplementation().readFile() without validation. An attacker could craft a sed edit command with a relative path like ../../etc/passwd to read arbitrary files.

The component reads and writes files based on user-controlled filePath without verifying the user has permission to access or modify that file. The _simulatedSedEdit bypass mechanism could be exploited.

GrowthBook analytics client keys are hardcoded in source code: 'sdk-yZQvlplybuXjYh6L', 'sdk-xRVcrliHIlrg4og4', 'sdk-zAZezfDKGoZuXXKe'. These keys can be extracted by code analysis and used to identify the application instance and potentially abuse GrowthBook analytics services.

allowUnsandboxedCommands defaults to true, enabling commands to bypass sandbox via dangerouslyDisableSandbox parameter. This defeats the primary security control when sandboxing is enabled.

The inbox poller runs every 1 second (INBOX_POLL_INTERVAL_MS = 1000) without any rate limiting on the poll function itself. An attacker could trigger this hook repeatedly through state changes, causing continuous mailbox reads and message processing that could exhaust CPU and network resources.

Remote session configuration including sessionId is logged in debug mode, potentially exposing user session identifiers

Skill improvement survey logs skill names in analytics events (line 45, 62). Skill names may contain PII or proprietary business information that users expect to remain private. No consent tracking or data minimization applied.

The stripTrailing function contains a while loop that iterates through characters without a maximum iteration count. Malicious input with extremely long trailing character sequences could cause CPU exhaustion through excessive iterations.

The wrapWithOsc8Link function accepts a url parameter and embeds it directly into terminal escape sequences without any validation. An attacker could inject malicious URLs or escape sequences that could be interpreted by terminal emulators.

setClipboard() accepts arbitrary text and writes it to clipboard via OSC 52 without sanitization. Malicious content could be injected into clipboard and potentially executed when pasted into vulnerable applications. The function doesn't validate or sanitize the input text before base64 encoding.

The SELECT_MEMORIES_SYSTEM_PROMPT is concatenated with user query and available memories without structural separation. An attacker could craft a query that contains adversarial instructions to manipulate memory selection, potentially causing the LLM to retrieve malicious or irrelevant memories based on injected commands rather than actual relevance.

The respondToPermissionRequest method sends permission responses to CCR without validating the requestId format or checking for replay attacks. The pendingPermissionRequests Map is cleared after sending response, but there's no rate limiting or duplicate request detection.

The extractSessionMemory hook runs as a forked agent with token usage tracking, but there is no rate limiting on how frequently extractions can be triggered. An attacker could force many extractions in quick succession, driving up token costs and resource usage.

The logEventTo1PAsync function calls getCoreUserData(true) which may include PII (accountUuid, organizationUuid, etc.) and logs it in event attributes. No consent tracking or user opt-out mechanism is present for this data collection.

logGrowthBookExperimentTo1P logs userAttributes including sessionId and userAttributes JSON string, which may contain PII. No consent tracking for this data collection.

Event logging exporter sends batches to external API without per-tenant or per-user spend caps. Adversarial traffic could trigger massive event logging volumes, driving up storage and API costs without budget protection.

JSON.parse() is called on CLAUDE_INTERNAL_FC_OVERRIDES env var without depth limiting. Maliciously crafted deeply nested JSON can cause stack overflow or memory exhaustion.

**Perspective 1:** extractToolInputForTelemetry truncates tool inputs but doesn't redact sensitive patterns like API keys, tokens, or credentials. The truncation at 512 chars may still expose sensitive information in the first portion. **Perspective 2:** Tool input telemetry collection may capture sensitive user inputs including API keys, credentials, or proprietary data before sanitization.

**Perspective 1:** getFileExtensionsFromBashCommand extracts file extensions from bash commands for analytics. While it limits extension length, it still logs file paths that could contain sensitive information (API keys, PII, internal paths) in the command arguments. **Perspective 2:** File extension extraction logic could expose sensitive filenames in analytics metadata, potentially revealing internal file paths or project structure.

uploadSessionFiles() processes file uploads in parallel with concurrency limit of 5, but no maximum array size validation. Each file upload is a billable operation. Attackers could submit thousands of files to drain budget.

The uploadSessionFiles endpoint accepts an unbounded array of files without size limits. Each file upload triggers storage costs and potential downstream processing costs.

The `findLastUuid` function and `fetchSessionLogsFromUrl` function expose internal session management details including UUID handling, session state recovery logic, and concurrent modification detection. Error messages reveal internal state tracking mechanisms.

Telemetry initialization error logs complete error details including potential sensitive context. This could leak internal state, configuration, or request information to analytics systems.

Diagnostic tracking service stores file URIs and diagnostic data in baseline maps without encryption at rest. The service tracks which files are opened, edited, and their diagnostic states across sessions. No consent tracking or data minimization for file path collection.

The opener() function concatenates existingMemories directly into the extraction agent's system prompt without provenance filtering or sanitization. If memories are user-submittable or editable, adversarial memories can inject instructions that the extraction agent will execute, enabling RAG poisoning attacks where poisoned content influences memory extraction decisions.

Memory extraction prompt concatenates existing memories directly into the system prompt without provenance filtering or delimiters. Untrusted or poisoned memories can be injected into the LLM context, causing the extraction process to generate compromised or adversarial memory entries.

The OAuth polling mechanism in performMCPOAuthFlow can hold connections open for up to 5 minutes (5 * 60 * 1000ms) without rate limiting. An attacker could trigger repeated authentication flows, exhausting server resources and connection pools.

The getChannelAllowlist function fetches the allowlist from GrowthBook feature flags without any integrity verification. An attacker could potentially inject malicious channel plugins into the allowlist, enabling unauthorized MCP server connections.

The gateChannelServer function returns a ChannelGateResult with a 'reason' field that exposes detailed internal policy checks including subscription type validation, org policy configuration, session allowlist state, and marketplace verification details. An attacker can fingerprint the authentication flow, subscription tiers, and policy enforcement mechanisms.

The `getClaudeAIOAuthTokens()` function is memoized for session lifetime but never checks token expiration. Tokens may be used after expiry, and there's no refresh mechanism. This creates stale token vulnerabilities where expired credentials remain valid in cache.

The runElicitationResultHooks function executes user-defined hooks that may trigger expensive downstream operations. No rate limiting or per-tenant spend caps on hook execution. Malicious hooks could trigger repeated LLM calls, file operations, or external API calls without cost controls.

The skipBrowserOpen option allows callers to bypass the browser-based OAuth flow entirely. When set to true, both manual and automatic URLs are passed to the caller who decides how to handle them. This could allow attackers to intercept the OAuth flow if the caller doesn't properly secure the URLs.

The token exchange at line 100-103 uses the authorization code directly without validating it against the expected state parameter. This could allow replay attacks where an attacker captures and reuses an authorization code.

Remote managed settings are fetched from an external API without cryptographic integrity verification. The service uses checksum-based HTTP caching but does not verify the authenticity of the settings payload. An attacker controlling the API endpoint could inject malicious configuration settings.

fetchWithRetry retries up to DEFAULT_MAX_RETRIES + 1 times (6 attempts) with exponential backoff but no maximum total spend cap. If the API endpoint is metered (e.g., per-request billing), this creates a 6x cost multiplier on failed requests with no budget tripwire.

Authentication errors (401, 403) skip retries with skipRetry: true, but this could allow attackers to probe auth boundaries. The error message 'Not authorized for remote settings' is generic but the fail-open behavior combined with the retry logic could enable enumeration attacks.

Remote managed settings are persisted to disk at line 445 with no explicit time-to-live (TTL) or automatic deletion mechanism. Settings may contain enterprise policy configurations that could include user-specific data. GDPR Article 17 (right to erasure) requires mechanisms to delete personal data upon request.

The lastAssistantText from user conversation is directly concatenated into the tool use summary prompt. An attacker could craft assistant messages containing adversarial instructions that manipulate the Haiku model to generate misleading summaries, potentially hiding malicious tool usage patterns.

VOICE_STREAM_BASE_URL is used directly in URL construction without validation. An attacker could inject malicious WebSocket endpoints via environment variable override, potentially enabling SSRF or connecting to attacker-controlled servers.

The SKILLIFY_PROMPT template directly interpolates user-supplied args and userMessages without sanitization. An attacker could inject malicious content into the prompt that gets sent to the LLM, potentially causing prompt injection attacks.

The filterToolsForAgent function allows certain tools for agents with minimal validation. The permissionMode checks may not properly enforce isolation between agents, potentially allowing privilege escalation if an agent gains access to tools it shouldn't have.

VERIFICATION_AGENT runs builds, tests, linters, and checks that may trigger expensive LLM API calls, GPU compute, or vector database queries. No rate limiting, no authentication checks on verification endpoint. Could be abused to exhaust GPU/LLM budgets.

Verification agent operations trigger LLM inference without rate limiting. Unlimited verification requests could drain LLM API budgets.

The buildForkedMessages function creates child agent messages without token budget limits. Forked agents inherit the parent's tool pool and can execute unlimited tool calls, potentially triggering unbounded LLM token consumption and API costs.

The loadAgentsDir function loads markdown files from 'agents' subdirectory using loadMarkdownFilesForSubdir. If the directory resolution is not properly sandboxed, an attacker could potentially load arbitrary files or inject malicious agent definitions.

Agent execution connects to MCP servers defined in agent frontmatter without verifying server integrity or validating that the server configuration hasn't been tampered with. Malicious agents could inject harmful MCP servers.

The sed edit parser accepts user-provided file paths and regex patterns without validation. A malicious user could craft a sed command with path traversal or shell metacharacters that execute arbitrary commands when the command is passed to the shell.

The 'dangerouslyDisableSandbox' flag is exposed in the BashTool input schema (line 127), allowing user-controlled input to bypass sandbox protections. This flag can be set to true to override sandbox mode and run commands without sandboxing, creating a direct path to command injection if the flag validation is insufficient.

The BashTool accepts arbitrary user-provided commands via the 'command' field and executes them through shell (line 1046). While there's permission checking and sandboxing, the combination of user-controlled command execution with a 'dangerouslyDisableSandbox' flag creates an attack surface where malicious commands could bypass security controls.

The BashTool executes shell commands from user input without validating command length or complexity. Commands can be arbitrarily long and complex, potentially causing resource exhaustion through CPU-intensive operations, large output generation, or infinite loops.

The BashTool accepts a 'dangerouslyDisableSandbox' parameter that can be set to true to override sandbox mode and run commands without sandboxing. This creates a potential privilege escalation vector where an attacker could exploit this to execute arbitrary code without isolation controls.

The shell command execution allows prefix injection through user input. Attackers can prepend shell commands using special characters or command chaining operators.

logClassifierResultForAnts function logs the full bash command including code and filepaths to analytics. Even though marked ANT-ONLY, this captures potentially sensitive code, credentials, or file paths that could be exposed in analytics systems.

The sed expression parser accepts user-provided patterns and replacements without proper escaping. While it validates flags, the pattern and replacement strings could contain shell metacharacters that, when executed via sed -i, could lead to command injection if the sed command is later used in a shell context.

TaskOutputTool passes task output to BashToolResultMessage with dangerouslyDisableSandbox: true, allowing user-controlled task output to bypass sandbox restrictions. This creates an injection vector where malicious commands in task output execute unsandboxed.

UNC paths (\\\\ or //) are skipped for filesystem operations to prevent NTLM credential leaks, but this check only validates the absolutePath. If a user provides a relative path that resolves to a UNC path after expandPath(), the bypass check fails. The validation should check the resolved path, not the input.

PowerShellTool executes shell commands without validating input size or rate limiting. An attacker can submit arbitrarily large commands or trigger resource-intensive operations that exhaust CPU/memory.

Background tasks can be spawned without rate limiting. An attacker could trigger many concurrent background tasks, exhausting system resources.

Shell command prefix injection possible through CLAUDE CODE SHELL PREFIX mechanism. User input can be prepended with malicious commands before execution.

The executeRemoteSkill function loads skills from external URLs (AKI/GCS) without verifying cryptographic signatures or tracking provenance. Skills are fetched and injected directly into the conversation without any integrity verification, allowing potential supply chain attacks where malicious actors could inject harmful code into the skill repository.

TaskOutputTool renders bash output with dangerouslyDisableSandbox: true, bypassing sandbox protections for user-controlled task output. This allows potentially malicious commands from background tasks to execute without isolation.

WebSearchTool processes multiple search queries in parallel via Promise.all without limiting total concurrent searches or total query count. Could be exploited to trigger massive parallel API calls.

**Perspective 1:** The baseUrl is used directly in fetch() without validating it has a valid HTTP/HTTPS scheme. An attacker could supply a malformed URL that might be interpreted differently by different browsers or proxies. **Perspective 2:** The fetch request doesn't validate response content length, potentially allowing large response bodies to be consumed.

**Perspective 1:** The autoUpdateMarketplacesAndPluginsInBackground function updates plugins without cryptographic verification. This could allow supply chain attacks through compromised plugin updates. **Perspective 2:** The cleanupNpmCacheForAnthropicPackages function removes npm cache without verifying package integrity. This could allow poisoned packages to be restored in future sessions.

Shell snapshot creation logs detailed error information including error codes, signals, working directory, Claude home directory, and full snapshot script content. This could reveal internal paths, environment configuration, and execution context to attackers who can trigger snapshot creation failures.

AWS credentials are cached and reused without validation of their freshness or revocation status. The cachedCredentials are stored and used directly without checking if they've been rotated or revoked by the user.

createBufferedWriter() defaults maxBufferBytes to Infinity, allowing unlimited memory accumulation before flush. An attacker could trigger rapid writes that fill available memory, causing OOM conditions. The deferred flush mechanism compounds this by batching writes without size constraints.

Claude Desktop integration reads configuration files from user-provided paths without proper validation. The path resolution could access unintended files or directories.

Chrome native host manifest is written to disk without verifying the manifest content's integrity or authenticity. The manifest contains allowed_origins and path fields that could be tampered with to redirect browser extension communication to malicious endpoints.

Default cleanup period is 30 days (DEFAULT_CLEANUP_PERIOD_DAYS = 30), but SOC 2 Type II requires audit logs to be retained for minimum 1 year. The cleanup function removes all logs including error logs and debug logs without distinguishing between compliance-critical audit trails and temporary debug data.

Audit logs are not retained for the required minimum period. SOC 2 Type II requires audit logs to be retained for at least 12 months. The cleanup logic at line 12 does not enforce minimum retention periods, allowing logs to be deleted before compliance requirements are met.

The ensureOsPermissions method checks accessibility and screen recording permissions via TCC. These checks could be bypassed or exploited to gain unauthorized system access. The adapter is a singleton that persists for the process lifetime.

Computer Use MCP tools (mcp__computer-use__*) are added to allowedTools array, bypassing normal permission prompts. The package's request_access handles approval for the whole session, but this creates a single point of failure - if the MCP layer is compromised, all computer use capabilities are available.

The handleDeepLinkUri function parses URIs without comprehensive input validation. Malformed or malicious URIs could lead to command injection or unauthorized access to local resources.

Protocol handler registration creates .app bundles and desktop files without verifying the integrity of the handler binary or checking for tampering. A malicious actor could register a compromised protocol handler.

execSyncWithDefaults_DEPRECATED accepts arbitrary command strings without validation or sanitization. This creates potential command injection vulnerabilities when user input is passed to the command parameter.

The fetchFastModeStatus() function makes HTTP requests to an external API endpoint without verifying response integrity, implementing timeout enforcement, or validating the authenticity of the response. This creates a potential man-in-the-middle attack vector where an attacker could intercept and modify the fast mode status response.

FINGERPRINT_SALT is hardcoded as a string literal in source code. This salt is used to compute user message fingerprints for attribution. Hardcoding cryptographic salts in source code is a security and privacy risk - the salt should be managed securely and not exposed in source control.

The query loop uses for-await without guaranteed termination. If the underlying stream doesn't properly close or sends excessive messages, this could cause indefinite CPU usage and memory growth from accumulating outputMessages array.

The regex /^(?:https?|ssh)://(?:[^@]+@)?([^/]+)/(.+?)(?:\.git)?$/ uses a non-greedy match (.+?) with potential for catastrophic backtracking on malicious inputs with many path segments and dots.

The extractGlobBaseDirectory function processes filePattern input without validating that the resolved path stays within intended boundaries. An attacker could craft patterns like '../etc/passwd' or absolute paths to access files outside the intended search directory.

The performHeapDump function captures heap snapshots and writes diagnostics, but there is no maximum frequency limit or budget circuit breaker. An attacker could trigger repeated heap dumps to exhaust resources and drive up costs through the underlying infrastructure.

The updateWatchPaths function accepts watchPaths from hook execution results without validation. Hooks could return arbitrary paths that could be used for path traversal attacks or to watch sensitive system directories.

Hooks are shell commands executed from .claude/settings.json without cryptographic integrity checks. While workspace trust is enforced, there's no verification that hook configurations haven't been tampered with or that hooks come from verified sources.

The `formatRecentMessages` function extracts user/assistant message content and concatenates it directly into the LLM prompt at line 61-62. The system prompt and user-controlled message content share the same string without clear delimiters or role markers, enabling prompt injection via adversarial user messages.

The `applySkillImprovement` function accepts `skillName` and reads from a file path derived from user input (`join(getCwd(), '.claude', 'skills', skillName, 'SKILL.md')`) without validation. This allows path traversal or injection of malicious content into the LLM's context.

Skill improvement events log _PROTO_skill_name which routes to privileged PII column. User preferences, corrections, and skill definitions are collected and logged without explicit consent tracking. The system tracks user behavior patterns and preferences during skill execution.

The `ssrfError` function exposes the complete list of blocked IP ranges (private, link-local, cloud metadata) in error messages. This helps attackers understand the security boundaries and potentially craft bypass attempts.

Image resizing/compression pipeline processes images without per-user or per-tenant rate limits. Adversarial users could trigger thousands of image processing operations, each consuming compute resources and potentially billing for GPU/TPU usage.

SessionInfo type exposes customTitle, gitBranch, cwd, tag, createdAt, and fileSize fields. Attackers can enumerate session directories, identify working directories, and fingerprint git repositories.

The listSessionsImpl function enumerates session files from project directories and worktrees. If worktree path resolution is vulnerable to path traversal, an attacker could potentially read session files from unintended locations.

The `--hard-fail` flag (line 205) causes the application to exit with code 1 on any error logged via `logError()`. This can cause unexpected crashes during normal operation and may expose internal error details to users.

Cached AWS credentials are stored in clientConfig and reused across requests. If these credentials are compromised, they remain valid until the cache is cleared. No expiration or rotation mechanism is implemented.

The findFirstUnescapedChar function uses a linear scan with backslash counting that could be exploited with carefully crafted input containing many backslashes followed by the target character, causing O(n²) behavior in worst-case scenarios.

The findLastUnescapedChar function has the same vulnerability pattern as findFirstUnescapedChar, scanning backwards with backslash counting that could cause catastrophic backtracking.

The resolveDependencyClosure function performs a DFS walk through plugin dependencies without limiting the maximum depth or number of plugins processed. An attacker could create a plugin with many transitive dependencies to exhaust CPU and memory during installation.

**Perspective 1:** The installPluginsForHeadless() function installs plugins from marketplaces without any cryptographic verification of plugin sources, no provenance tracking for installed artifacts, and no integrity verification of downloaded plugin ZIPs. This creates a supply chain attack vector where malicious plugins could be injected into the system. **Perspective 2:** The plugin installation process does not generate Software Bill of Materials (SBOM) for installed plugins. This prevents supply chain visibility into plugin dependencies and makes vulnerability management difficult. **Perspective 3:** The plugin installation process lacks deterministic build controls. Plugin installations may vary between runs, making security auditing and reproducibility verification impossible.

The detectAndUninstallDelistedPlugins() call at line 54 relies on a security.json fetch that has been removed from the codebase. This creates a gap where delisted malicious plugins may not be detected and uninstalled in headless mode.

**Perspective 1:** Plugin install counts fetched from GitHub without cryptographic verification or provenance tracking. The axios GET request to 'https://raw.githubusercontent.com/anthropics/claude-plugins-official/[branch-redacted]' has no signature verification, making it vulnerable to supply chain attacks where an attacker could inject false install counts or malicious data. **Perspective 2:** GitHub token exposure risk - if the stats repository requires authentication or if the public repo is compromised, an attacker could modify the install counts data. No rate limiting or authentication headers are implemented.

**Perspective 1:** Plugin installation metadata is managed without cryptographic verification of plugin sources. The installed_plugins.json file tracks plugin installations but there is no signature verification or provenance tracking for downloaded artifacts. This allows potentially malicious plugins to be installed without verification. **Perspective 2:** Plugin installations do not enforce build reproducibility controls. Plugins could be built with different configurations at different times, leading to non-deterministic behavior and potential security gaps.

The addInstalledPlugin function accepts plugin metadata without verifying the source against an official registry. Plugins can be installed from unverified sources, creating a supply chain attack vector where malicious code could be injected via plugin installation.

**Perspective 1:** Plugins are loaded from user-provided directories (plugin.commandsPath, plugin.skillsPath) without signature verification, checksum validation, or provenance tracking. Malicious actors could distribute tampered plugins that appear legitimate. **Perspective 2:** The codebase does not generate SBOMs for plugins or dependencies. This prevents supply chain visibility, vulnerability tracking, and compliance with security standards like NIST SP 800-161. **Perspective 3:** Plugin dependencies are loaded without integrity checks (hash verification, checksum validation). This allows supply chain attacks where dependencies could be replaced with malicious versions. **Perspective 4:** Plugin loading uses memoization (lodash-es/memoize) which could cache malicious plugin states. If a plugin is compromised after initial load, the cached state persists until cache.clear() is called. **Perspective 5:** Plugin dependencies are loaded without explicit version pinning or semantic version constraints, potentially allowing supply chain attacks via dependency updates.

Plugins are loaded from inlinePlugins and enabled plugins without verifying they come from a trusted source registry. The code accepts any plugin path without validating against known-good sources or checking for dependency confusion attacks.

**Perspective 1:** Plugin hooks are registered from loaded plugins without any cryptographic signature verification or provenance tracking. Hooks can be injected from untrusted sources without integrity validation. **Perspective 2:** No Software Bill of Materials (SBOM) is generated for plugin artifacts or hook configurations, making dependency tracking and vulnerability assessment impossible.

Managed plugins are retrieved from policy settings but there is no cryptographic verification or provenance tracking for plugin artifacts. This allows unverified plugins to be loaded into the execution environment.

The cacheMarketplaceFromUrl function accepts arbitrary URLs without validation. An attacker could provide a URL pointing to an internal service (e.g., http://localhost:8080/admin) to perform SSRF attacks and access internal resources.

MCP server URLs are constructed from user input without proper validation. The resolvePluginMcpEnvironment function expands ${VAR} patterns and uses URLs directly in server configuration, enabling SSRF attacks.

The official marketplace fetch from GCS downloads and extracts plugins without verifying digital signatures or tracking artifact provenance. Plugins are installed based solely on SHA comparison against a sentinel file, with no cryptographic verification of the plugin contents themselves. This creates a supply chain attack vector where malicious plugins could be injected into the GCS bucket.

The unzipFile function extracts files from a zip archive without any size limits on the extracted content. A malicious zip file with small headers but massive extracted content (decompression bomb) could exhaust disk space and memory during extraction.

autoUpdateMarketplacesAndPluginsInBackground() refreshes marketplaces and updates plugins without rate limiting. An attacker could trigger repeated marketplace refreshes via plugin operations, causing repeated git operations and API calls against paid marketplace infrastructure.

Comment indicates security.json fetch was removed (see #25447) - ~29.5M/week GitHub hits for UI reason/text only. This means plugins that have been removed from marketplaces are not being detected and auto-uninstalled, allowing potentially malicious or deprecated plugins to remain installed.

getPluginSeedDirs() splits CLAUDE_CODE_PLUGIN_SEED_DIR by delimiter and maps through expandTilde, but doesn't validate that resulting paths are within expected boundaries. Could allow path traversal to read arbitrary files.

validateGitUrl() accepts http:// URLs alongside https://, creating potential SSRF vulnerability when plugins are installed from user-controlled git URLs. HTTP traffic is not encrypted and can be intercepted.

reconciler.ts accepts marketplace sources from settings without verifying them against a trusted registry. The normalizeSource function resolves paths without integrity checks, allowing supply chain attacks through malicious marketplace configurations.

Plugin zip cache extraction allows file path traversal through user-controlled paths in plugin manifests. The path validation does not properly sanitize directory traversal sequences.

checkEndpoints() makes HTTP requests to oauthConfig.BASE_API_URL and tokenUrl.origin without validating the host. An attacker with control over oauthConfig could redirect requests to internal services (127.0.0.1, localhost, internal IPs) for reconnaissance or data exfiltration.

The readEditContext function accepts user-provided file paths without validation, allowing potential path traversal attacks. The function opens files at arbitrary paths and scans them for context, which could be exploited to access files outside intended directories.

The resolveSandboxFilesystemPath function accepts path patterns without full validation. Paths like '/Users/foo/.cargo' could be misinterpreted as settings-relative rather than absolute, potentially allowing access to unintended filesystem locations.

The sandbox configuration has `allowUnsandboxedCommands` defaulting to `true` (line 358). This allows potentially dangerous commands to bypass sandbox restrictions, enabling code execution attacks through the Bash tool.

On non-macOS platforms (Linux, Windows), the application falls back to plaintext storage instead of using platform-native secure storage. This means credentials are stored unencrypted on all platforms except macOS, creating inconsistent security posture.

macOS keychain storage falls back to plaintext storage if keychain operations fail. This creates a security gap where credentials may be stored unencrypted during keychain failures or on systems where keychain is unavailable.

The storagePath is hardcoded to ~/.claude/.credentials.json with no encryption. Even with chmod 0o600, credentials are readable by any process that can access the user's home directory. This violates defense-in-depth principles for credential storage.

Secure storage implementation stores credentials in plaintext JSON file with only file permissions (0o600) protection. No cryptographic encryption or signing of stored data. If file permissions are bypassed or system is compromised, all credentials are exposed.

File is set to 0o600 (owner read/write only), but since the content is unencrypted plaintext, any user with read access to the filesystem can extract credentials. Combined with plaintext storage, this provides false sense of security.

Session environment scripts are written to disk at line 12 without any encryption. These files contain environment variables and hook scripts that may include sensitive configuration, API keys, or authentication tokens. GDPR requires appropriate security measures for personal data at rest.

The readFile operations in getSessionEnvironmentScript() read hook files without any size limits. A malicious hook file could be extremely large, causing memory exhaustion when loaded into the session environment script.

Session IDs for JSONL file paths are generated using randomUUID() instead of validating against an existing session. This allows attackers to create arbitrary session IDs by providing any .jsonl filename, potentially enabling session hijacking or replay attacks.

Any URL is accepted as a valid session identifier and a random UUID is generated. This bypasses session validation entirely - an attacker could craft arbitrary ingress URLs to create fake sessions, enabling session fixation or unauthorized session creation.

runSideQuestion() calls runForkedAgent() with maxTurns: 1 but no token limit. The wrappedQuestion can be arbitrarily large, and the forked agent can consume unlimited tokens. Missing per-session spend cap.

TaskOutput writes command output to disk files (task output paths) without encryption. File mode stdout/stderr are persisted to disk via DiskTaskOutput. This violates encryption-at-rest requirements for sensitive data.

TaskOutput stores output files indefinitely without explicit retention policy. Files are only deleted via explicit deleteOutputFile() call or when outputFileRedundant becomes true. No automatic cleanup based on age or size limits.

The BigQueryMetricsExporter class contains a hardcoded internal API endpoint URL (https://api.anthropic.com/api/claude_code/metrics) that reveals internal infrastructure details. This endpoint is used for metrics collection and could help attackers identify the organization's internal API structure.

The telemetry shutdown uses Promise.race with a timeout, but the timeout value can be configured via environment variable. If set to a very high value, this could block process exit indefinitely, preventing graceful shutdown and resource cleanup.

Perfetto tracing captures detailed internal agent hierarchy (parent-child relationships), API request metadata (TTFT, TTLT, prompt length, cache stats, message IDs), tool execution details, and user input waiting times. When enabled via CLAUDE_CODE_PERFETTO_TRACE, this exposes the complete internal architecture and conversation flow to anyone who can access the trace file.

User email address is collected and included in telemetry attributes without redaction or consent mechanism. If logs are accessible to unauthorized parties, this exposes PII.

Email addresses are being logged in telemetry attributes, exposing user PII. This violates privacy requirements and could lead to user identification and tracking.

User account UUID is logged in telemetry attributes. This is a persistent identifier that can be used to track users across sessions and potentially link to other systems.

Account UUIDs are being logged in telemetry attributes, which can be used to identify and track specific users across sessions. This is PII that should not be logged.

The getOauthAccountInfo() function returns organizationUuid, emailAddress, and accountUuid which are then added to telemetry attributes. These sensitive identifiers are being sent to analytics/telemetry services without redaction, potentially exposing user identity and organizational information.

User email address is being collected and sent to telemetry services via attributes['user.email']. This is PII that should not be transmitted without explicit user consent and proper data governance.

OAuth account UUID is being sent to telemetry services via attributes['user.account_uuid'] and attributes['user.account_id']. This unique identifier can be used to track and identify specific users across sessions.

**Perspective 1:** The teleportToRemote function uploads git bundles via createAndUploadGitBundle without any cryptographic verification or signing. This allows potential tampering with repository contents during teleport session creation. **Perspective 2:** Teleport sessions are created and resumed without provenance tracking. This prevents audit trails of session creation, modification, and execution.

The pollRemoteSessionEvents function polls remote session events with a MAX_EVENT_PAGES cap of 50, but there is no rate limiting on how frequently this polling can be triggered. Continuous polling without backoff could exhaust API rate limits and cause resource exhaustion.

Session logs are fetched via getTeleportEvents and getSessionLogsViaOAuth without any integrity verification. An attacker could potentially inject malicious messages into resumed sessions.

The `fetchSession()` function has a 15-second timeout but no session inactivity timeout enforcement. Sessions remain active indefinitely without expiration, allowing attackers to hijack abandoned sessions. No idle timeout or maximum session age validation.

The bundlePath from generateTempFilePath() is used directly in execFileNoThrowWithCwd() without validation. While generateTempFilePath() likely creates safe paths, there's no defense-in-depth validation before use.

The segmentTextByHighlights function tokenizes user-supplied text without any size validation. An attacker could supply extremely large text inputs causing O(n) memory allocation and CPU consumption during tokenization, potentially exhausting process resources.

The whichNodeAsync function uses execa with shell: true and passes user-controlled command parameter directly to the shell. An attacker could inject additional commands via shell metacharacters (e.g., `; rm -rf /`, `| cat /etc/passwd`, `&& whoami`).

The whichNodeSync function uses execSync_DEPRECATED with shell: true and user-controlled command. Same injection vector as async version - shell metacharacters can execute arbitrary commands.

logForDebugging calls at line 325 and throughout the file expose internal worktree session state, paths, and git operations. These logs may be captured and reveal repository structure, session identifiers, and internal implementation details to anyone with access to logs.

YAML parsing via Bun.YAML or yaml package without validating input structure. YAML parsers can be exploited for arbitrary object creation and prototype pollution attacks if untrusted input is parsed.

Bridge session tokens and URLs are resolved without enforcing artifact signing or verification. The bridge configuration allows override tokens and URLs that could be used to redirect to untrusted endpoints without cryptographic verification.

Debug logging includes session_id and uuid fields in control_response messages. This can be used to track and correlate user sessions.

BridgePointerSchema requires sessionId as a string but provides no validation to ensure it contains sufficient entropy. If session IDs are generated elsewhere with weak randomness (e.g., Math.random(), incrementing counters), they could be predictable.

BridgePointerSchema requires environmentId as a string but provides no validation to ensure it contains sufficient entropy. Predictable environment IDs could allow attackers to guess or enumerate bridge sessions.

The bridge pointer file (bridge-pointer.json) contains sessionId and environmentId that persist across crashes. If this file is accessible via filesystem enumeration, attackers can identify active remote control sessions and correlate them with specific environments.

The bridge status state machine has no idle timeout mechanism. A disconnected bridge connection remains in 'idle' or 'reconnecting' state indefinitely. This could allow attackers to maintain bridge access after legitimate disconnection.

updateFailedStatus displays the raw error string to users via chalk.red(error), potentially exposing internal error details or stack traces.

The baseUrl parameter is used directly in URL construction without validation. Could allow protocol injection or malformed URLs.

The title parameter is used directly in API request without length or content validation.

The createBridgeSession function accepts an events array without size validation. A malicious caller could send thousands of events, creating a large request body that could exhaust server-side resources or cause timeout issues.

Bridge configuration values are fetched from GrowthBook feature flags without build reproducibility controls. Configuration changes could introduce non-deterministic behavior across builds. No version pinning or reproducible build constraints are enforced.

getBridgeAccessToken() is called to retrieve OAuth tokens for file downloads. If error handling exposes token values or the token retrieval process reveals authentication mechanisms, it could leak internal OAuth configuration.

logForDebugging calls expose bridge initialization state including OAuth token expiration, session titles, git context (branch, remote URL), and worker type. These logs could be captured or leaked.

The OAuth token refresh happens in background with fire-and-forget pattern. If refresh fails, the stale token remains in use until next explicit refresh, potentially allowing continued access with compromised credentials.

JWT tokens are decoded without verifying the signature. While this is a standard practice for certain use cases, it could be exploited if tokens are tampered with before reaching this code.

JWT expiry claims are decoded without signature verification. Malicious actors could craft tokens with arbitrary expiry times if tokens are not properly validated elsewhere.

Comment references 'Server PR #292605 (renamed in #293280)' which appears to be a hallucinated or non-existent PR reference. This type of specific reference without verifiable context suggests AI-generated documentation.

The createV2ReplTransport function creates WebSocket/SSE connections using a JWT token, but the authentication validation relies on the token being provided via getAuthToken callback. There's no fallback validation if the token is invalid or expired before the connection is established.

The sessionUrl parameter is used directly in URL construction without validation. Malformed or malicious URLs could cause issues.

enrollTrustedDevice() makes POST requests to external API without rate limiting or per-tenant spend caps. Could be abused to trigger expensive downstream processing or API charges.

Bridge enrollment operations lack per-tenant budget caps. Each enrollment could trigger expensive authentication flows and downstream processing.

BridgeConfig type includes apiBaseUrl and sessionIngressUrl fields that are transmitted to the client. If these URLs contain internal infrastructure information, they could aid reconnaissance attacks against the backend infrastructure.

The workSecret module handles session ingress tokens and API base URLs without any provenance tracking or verification of the secret's origin. This could allow session hijacking or unauthorized access if secrets are compromised.

Error messages expose internal implementation details including version information and data structure validation logic.

Error messages may expose network addresses and hostnames that could be used for targeted attacks.

The content parameter in prependUserMessage() has no length validation, allowing arbitrarily large user messages that could exhaust memory or cause denial of service.

Duplicate control_response error logging includes request_id and toolUseID which could be used to correlate attacks or track tool usage patterns.

Hook callback error logging includes callbackId which could be used to enumerate active hooks or correlate hook invocations.

Connection error logging exposes full error message which may contain sensitive context

Stream read error logging exposes complete error message to diagnostics

**Perspective 1:** The SerialBatchEventUploader stores events indefinitely with no TTL or deletion policy. Sensitive events (user activities, tool usage) should have a defined retention period for GDPR compliance. **Perspective 2:** The T type parameter for batch events has no data classification mechanism. Events should be classified (PII, sensitive, public) to determine appropriate handling and retention.

WebSocket URL passed directly to constructor without validating scheme (ws:// vs wss://). Could allow downgrade attacks or connection to malicious endpoints.

The state coalescing logic may not correctly handle concurrent patch operations, potentially losing data when multiple patches arrive simultaneously.

The `writeEvent()` method sends POST requests to `/worker/events` without rate limiting. Rapid-fire event submission could overwhelm the server or enable DoS attacks. The SerialBatchEventUploader has queue size limits but no per-request rate limiting.

The `writeInternalEvent()` method sends POST requests to `/worker/internal-events` without rate limiting. Similar to client events, this could be exploited for DoS or resource exhaustion attacks.

The `reportDelivery()` method sends POST requests to `/worker/events/delivery` without rate limiting. High-frequency delivery updates could exhaust server resources.

The expandPath function is used before resolve(), but there's no explicit check for path traversal patterns like .. or encoded variants. The validation relies on stat() which may not catch all edge cases.

Error handling displays raw error messages to users without sanitization, potentially exposing internal error states and stack traces.

The customInstructions string from args is used without validation for length, encoding, or injection attempts.

Error handling re-throws specific error messages (ERROR_MESSAGE_NOT_ENOUGH_MESSAGES, ERROR_MESSAGE_INCOMPLETE_RESPONSE) that leak internal error classification

Generic error handling logs the full error object without sanitization, potentially exposing internal stack traces or error details

Error message includes the full error string which could leak internal implementation details or sensitive context

Comment claims isEnvTruthy() correctly handles DISABLE_COMPACT, but no test or verification exists for this specific use case.

Context usage data collection (messages, tools, agents, MCP servers) does not implement data classification. May collect and store sensitive user data without proper handling. Violates SOC 2 CC6.2 (data classification) and GDPR Article 5(1)(c).

Comment claims getIsNonInteractiveSession() is the correct check, but no enforcement or verification exists for this pattern.

Error handling in handleWrite displays full error message including potential file path information to end users via onDone callback. This could leak sensitive file paths or internal error details.

Effort level commands log the effort value setting in analytics events. This tracks user preference changes which could be used to build user profiles or identify individual users.

Login flow and API key changes are handled at line 8 without explicit user consent for logging authentication events. Authentication state changes may be tracked in analytics, potentially violating GDPR principles on data collection transparency.

The extra-usage command allows users to bypass usage limits without proper rate limiting or quota enforcement. This could lead to resource exhaustion.

The applyFastMode function checks if model switch is needed and updates state, but if multiple concurrent calls happen, the model switch logic may execute multiple times or with stale state, causing inconsistent behavior.

/init command analyzes entire codebase without file size limits or rate limiting. Could be exploited to trigger expensive file processing operations on large repositories.

Comment claims 'exit code is irrelevant — an unrecognized --version still means cmd exists' but this logic is flawed. Exit codes from spawnSync are meaningful for determining command availability.

Logout command has no audit trail. GDPR requires logging of access and modification events for personal data. Without logout audit logs, users cannot verify when their session was terminated, and admins cannot track unauthorized access attempts.

Logout removes local credentials but doesn't enforce a limit on concurrent active sessions. An attacker who obtains a session token can maintain access while the legitimate user logs out. No session count tracking or forced logout of other devices.

The args parameter is split and used to determine command actions without proper sanitization. Could allow command injection or parameter tampering.

The target parameter from user input is used directly in filter operations without validation. Could allow injection of malicious server names or bypass intended filtering.

OAuth refresh command is stubbed with isEnabled: () => false. This indicates missing or incomplete implementation of OAuth token refresh mechanisms, which could lead to credential expiration issues and lack of proper access token lifecycle management.

Error logging includes errorMessage(err) which may expose internal error details to users

Plugin reload command fetches user settings from remote endpoint without verifying the response's integrity or authenticity. Settings could be tampered with to inject malicious plugins.

The command exposes policy configuration details ('allow_remote_sessions') to users. This could leak internal policy information about what features are enabled.

The environment creation and token import functions do not include any mechanisms to ensure deterministic, reproducible builds. There's no version pinning, dependency locking, or build artifact signing to guarantee consistent deployments.

The remote setup command does not generate SBOM for session artifacts or track build reproducibility. Remote setup artifacts could be tampered with without detection.

GitHub tokens are imported for remote setup without verification of the token's provenance or the artifacts it will access. Token could be compromised or misconfigured.

Function accepts `args: string` parameter but immediately checks `!args || args.trim() === ''` - the parameter is accepted but the function doesn't use the original string value, only the trimmed version. This is a common AI pattern of accepting parameters without proper usage.

Error message 'This session is a swarm teammate. Teammate names are set by the team leader.' reveals internal session architecture and team leader concepts that could help attackers understand the system's organizational model.

Session names are saved and synced to bridge at line 28. Session names may contain personal identifiers or project names that could be considered PII under GDPR. No consent tracking or user notification mechanism exists for storing this data.

updateBridgeSessionTitle is called without rate limiting or authentication verification. If this endpoint triggers paid operations (session updates, metadata storage), repeated calls could accumulate costs without user consent or budget caps.

Error handling displays full error message to users via onDone callback. This could expose internal implementation details about ultrareview failures.

Error messages contain verbose internal state information that could be exploited for reconnaissance.

Error logging in QR code generation failure may include sensitive context about the URL or internal state.

Vim mode toggle command is registered but lacks explicit consent tracking for any telemetry that may be associated with this feature. Users may not be aware their editing mode preferences are being tracked.

Auto-update success/failure events log 'installationType' and 'wasMigrated' fields. This reveals system configuration details about how users installed the application, which could be used to identify or profile users.

OptionMap extends Map without size constraints. An attacker could create thousands of options, causing memory exhaustion and potential UI rendering DoS.

The desktop handoff calls gracefulShutdown(0, 'other') immediately after opening the desktop app. If the desktop app fails to connect, the CLI session is terminated abruptly, potentially leaving the user in an inconsistent state or allowing session state manipulation.

User-provided filename is used directly with path.join() without validation. An attacker could provide a path traversal sequence like '../../etc/passwd' which would be written outside the intended directory. The filename is only validated for .txt extension, not for path safety.

The hasMemoryFileRead function performs O(n) iteration through all messages on each render. With very large conversation histories, this could cause CPU exhaustion. The survey trigger also scans messages repeatedly.

The isSubsequence function uses O(n*m) naive matching that can be exploited with crafted inputs causing quadratic time complexity.

Configuration error dialog displays filePath which may reveal sensitive file locations

Error description field may contain internal validation error messages

Analytics events log session identifiers and user behavior data including 'tengu_session_search_toggled', 'tengu_session_tag_filter_changed', 'tengu_session_all_projects_toggled'. These events track user session activity which could be used to identify individual users.

Recent activity feed logs user session summaries with timestamps. This creates a record of user activity that could be used to track individual user behavior patterns over time.

The onChange handler logs 'tengu_mcp_dialog_choice' analytics events that include the server name being approved/rejected. This exposes which MCP servers are available and being considered by users, helping attackers map the MCP server ecosystem.

Comment references '#24180' as an issue number for 'turn50→turn99 RSS regression' which appears to be a hallucinated issue reference pattern.

Session ID is extracted from logs but there's no correlation ID tracking for audit purposes

The condition `if ("external" === 'ant')` is always false (string literal comparison). This dead code path appears to be scaffolding that was never properly implemented.

Session ID is displayed in the status UI, which could be used for session enumeration or tracking. While not directly exploitable, this leaks internal session identifiers to users.

The audioLevels array from useVoiceState is used directly without size limits. During voice recording, this could grow unbounded, consuming increasing memory and causing performance degradation over time.

Comment states 'eslint-disable @typescript-eslint/no-require-imports' but this rule doesn't exist in the codebase. The disable comment is meaningless and suggests AI-generated code that doesn't match actual project configuration.

The FuzzyPicker component renders all visible items in a single render cycle. If items.length is very large (thousands), this could cause memory pressure and slow rendering, especially with complex renderItem functions.

The handleAuthenticate function catches errors and displays err.message directly to users. This may leak internal implementation details about authentication failures, including potential error codes or internal state information.

MCP server names from user configuration are rendered in the MCP list panel and could be passed to LLM contexts without filtering for adversarial content or injection attempts.

getSessionIngressAuthToken() is called at line 28 to check authentication status. If error handling or logging exposes this value, it could leak internal session authentication tokens to users or attackers.

Error opening memory file displays the full error object which may contain internal implementation details or stack traces.

The sed edit permission request component uses simulated validation that doesn't match the actual sed command execution validation, creating a gap where user-controlled input could pass permission checks but fail at execution time.

The editablePrefix state (derived from user input) is passed to the classifier LLM call without sanitization. Malicious prefixes could be used to inject adversarial instructions into the classifier's context.

Classifier descriptions and review content are logged to analytics when users accept 'classifier-reviewed' option. This could capture sensitive code review content or security-related feedback.

pathFromToolUse() extracts file paths from tool inputs using tool.getPath() but doesn't validate the resulting path for traversal attacks or special characters before use.

User-provided feedback and instructions are logged to analytics via logEvent('tengu_accept_submitted') and logEvent('tengu_reject_submitted'). This could capture sensitive user intent or instructions about what operations to allow/deny.

The onSelect function logs network host patterns to analytics when USER_TYPE === 'ant'. This could expose internal network infrastructure details.

The host pattern from sandbox permission requests is logged when USER_TYPE === 'ant', exposing network connection information in debug logs.

The host pattern is logged again when the user cancels the permission dialog, creating duplicate log entries with sensitive network information.

The onCancel function logs network host patterns to analytics when USER_TYPE === 'ant'. This exposes internal network infrastructure details.

The SedEditPermissionRequest component handles file modifications but the audit trail for permission changes is incomplete. File edits are tracked via fileHistoryTrackEdit but there's no centralized audit log entry for the permission approval itself.

The SedEditPermissionRequest creates a simulated sed edit that applies changes directly without running the actual sed command (line 115). This bypasses sed's regex validation and could allow injection of malicious content through the sed syntax itself. The permission dialog shows a preview, but the actual execution path differs from what the user sees.

The component displays detailed sandbox configuration including excluded commands, filesystem restrictions, network restrictions, and Unix socket configurations. This reveals internal security policies and infrastructure details that could help attackers understand attack surface.

Sandbox dependencies check for tools (bubblewrap, socat, ripgrep) but do not verify container base image provenance or integrity. Malicious base images could compromise sandbox isolation.

Comment claims memoize(getUserContext) provides date stability, but no code demonstrates the cache busting scenario or verifies the fix.

process.env.USER_TYPE is used in a conditional without validation. While it's a build-time define, the check doesn't validate the value format.

GrowthBook analytics client keys are hardcoded as string literals in source code. These keys (sdk-yZQvlplybuXjYh6L, sdk-xRVcrliHIlrg4og4, sdk-zAZezfDKGoZuXXKe) are exposed in the repository and could be used to access analytics data or perform unauthorized operations.

Error messages may expose authentication tokens and credentials to users, potentially leaking sensitive information in logs or error output.

doInitializeTelemetry catches errors and logs them with errorMessage(error) which may include stack traces or sensitive context.

MCP server exposes tools without pinning dependency versions or using deterministic build flags. Non-deterministic builds could introduce supply chain vulnerabilities between deployments.

The file suggestions hook processes file lists without hard limits on the number of files or total size. A repository with hundreds of thousands of files could cause memory exhaustion during indexing or suggestion generation.

File paths from git ls-files and ripgrep are loaded into the file index without validation. An attacker could potentially supply paths outside the project root through git operations or configuration.

The generateFileSuggestions function processes user input (partialPath) and uses it to search the file system. While primarily a read operation, improper path validation could enable directory traversal attacks.

Installation messages are displayed to users and may contain paths or aliases that reveal user-specific information. These should be tracked with consent.

The useMcpConnectivityStatus hook tracks MCP server connections but doesn't verify authentication status before displaying connectivity information. This could expose connection details for unauthenticated or compromised servers.

The useMcpConnectivityStatus hook displays MCP server configuration details including server names and connection types in notifications. This could expose internal infrastructure details to users.

The hook calls getSubscriptionType() and hasClaudeAiBillingAccess() which could leak whether the user has team/enterprise subscription or billing access. This information helps attackers fingerprint the deployment environment and identify enterprise customers.

The handler races between bridge callbacks, channel callbacks, hooks, and classifier checks. The claim() guard prevents double-resolution but the race window could allow inconsistent state if one path fails after another succeeds.

The verification flow skips retrieving the API key from the helper before showing the trust dialog (skipRetrievingKeyFromApiKeyHelper: true). This could allow RCE via settings.json to be executed before user consent is obtained.

Line 23 contains `if (true && !isClaudeAISubscriber())` - the hardcoded `true` makes this condition always evaluate to `!isClaudeAISubscriber()`, bypassing any intended conditional gating. This could expose the subscription check to unintended execution paths and may indicate incomplete feature flag implementation.

useCommandQueue subscribes to command queue without rate limiting. Rapid queue mutations could trigger expensive downstream processing. No concurrency limits on queue operations.

useCommandQueue() subscribes to command queue without any maximum queue size. An attacker could flood the system with commands that accumulate in the queue, consuming memory and preventing legitimate commands from being processed.

The showDiff function logs 'tengu_ext_will_show_diff' and 'tengu_ext_diff_accepted' events. While file paths are passed to the function, it's unclear if they're included in the logged metadata, which could expose sensitive file locations.

useDynamicConfig() fetches config on every mount without debouncing or rate limiting. If used in high-frequency components, this could trigger excessive API calls to config services.

The eventData object is passed directly to logEvent without validation. Malicious input could potentially inject arbitrary data into logs, including log injection attacks.

User-controlled eventData is passed directly to logEvent without validation, creating potential log injection vulnerability where malicious input could be logged.

The useMainLoopModel hook does not generate SBOM for session artifacts. Model configuration changes could be tampered with without detection.

The installPluginAndNotify function installs plugins from marketplace without visible signature verification. An attacker could potentially inject malicious plugins if the marketplace is compromised.

User message content is logged in debug mode which could expose sensitive user input

WebSocket error handler logs full error.message which could leak internal error details or connection state

Cleanup error handling in useRemoteSession could expose internal state through error messages

useReplBridge handles OAuth authentication for bridge connections but doesn't validate token refresh mechanisms. Token expiration could lead to unauthorized access or authentication bypass.

Skill improvement survey error logging includes _PROTO_skill_name which could leak skill names to analytics systems.

The handleChange function calls clearCommandsCache() on every skill file change. Rapid file modifications could trigger excessive cache clearing operations, causing CPU spikes.

Task list management hook uses fs.watch and polling mechanisms without deterministic behavior guarantees. Build reproducibility is not enforced for task state persistence.

The fullAudioRef.current array accumulates all audio chunks during recording for silent-drop replay. Without a size limit, a 10-minute recording could consume ~20MB of memory. The buffer is never explicitly bounded.

Event emitter disables maxListeners warning by setting limit to 0. This could allow memory exhaustion attacks through unbounded event listener accumulation.

The error message 'Node is not in a tree with a FocusManager' reveals internal architecture details about the FocusManager system. Attackers could use this to understand the focus management implementation and potentially craft exploits targeting focus-related functionality.

The instances map stores all Ink render instances without any cleanup or size limits. An attacker could create many concurrent render instances to exhaust memory.

When applying styles to wrapped text, segment.hyperlink is used directly without validation. The URL could contain malicious content that bypasses the initial validation.

The isXtermJs() function checks TERM_PROGRAM environment variable and the setXtversionName() function processes terminal probe responses. These could leak terminal emulator type and version information to attackers who can trigger terminal initialization.

tabStatus() function accepts user-provided status text and emits it in OSC 21337 sequence without validation. The status field could contain malicious content that terminals might render. Only basic escaping of \ and ; is applied.

logCustomBindingsLoadedOncePerDay logs user_binding_count which reveals user customization behavior patterns. This could be considered PII as it tracks individual user configuration choices.

The loadKeybindings function parses user-provided keybindings.json without strict schema validation. Malformed or malicious JSON could cause unexpected behavior or information disclosure.

PathTraversalError messages include the full path being validated, which could help attackers understand the team memory directory structure.

Team memory system stores shared memories without explicit data classification or PII detection. Violates SOC 2 CC6.2 and potential HIPAA requirements if team memories contain PHI. Warning in code mentions avoiding API keys but no enforcement.

The error log includes the full error message which may contain sensitive context about the migration process, configuration state, or internal implementation details.

The color-diff module uses highlight.js via dynamic require() without verifying the integrity of the loaded module. This creates a supply chain attack vector where a compromised highlight.js version could execute malicious code.

The layoutNode recursive function doesn't have depth limits or iteration guards. A maliciously crafted flexbox tree with 10,000+ nested containers could cause stack overflow or severe performance degradation.

Built-in plugins are registered without explicit version pinning or integrity verification. Version drift could introduce unvetted changes to built-in functionality.

The connect() method creates a SessionsWebSocket with an access token callback, but there's no explicit validation that the token is valid before allowing the WebSocket upgrade. The SessionsWebSocket implementation (not in this diff) may lack proper auth header validation on the upgrade request.

The RemoteSessionManager receives a getAccessToken callback that returns session ingress tokens. These tokens are used for WebSocket connections and API calls. While the token is not hardcoded, ensure the callback implementation properly encrypts tokens at rest and uses HTTPS for all transmission.

The WebSocket URL is constructed from BASE_API_URL with session ID. If BASE_API_URL can be manipulated or if session IDs are predictable, this could lead to SSRF to internal services.

Error logging includes errorMessage(error) which may expose internal error details

The file imports randomUUID from 'crypto' which is the Node.js CSPRNG, but it's used to generate UUIDs for synthetic assistant messages in remote permission requests. While randomUUID is cryptographically secure, the synthetic messages are used in a security-sensitive context (permission requests) where predictable identifiers could potentially be exploited in replay attacks or message correlation attacks.

The speculation.ts file processes file paths through relative() but doesn't validate that the resolved path doesn't escape the working directory. An attacker could potentially use path traversal sequences to write files outside the intended overlay directory.

Session memory prompts can be triggered by user commands without token usage limits. The buildSessionMemoryUpdatePrompt function processes full conversation history which could be large, consuming LLM tokens without caps on how often this operation can be triggered.

The error response includes the full error message from the catch block, which may contain internal implementation details or sensitive context.

No Software Bill of Materials (SBOM) is generated for session artifacts or build outputs. Without SBOM, dependency vulnerabilities cannot be tracked through the build pipeline or session lifecycle.

The exporter sends events in batches with batchDelayMs defaulting to 100ms. An attacker could flood the system with events, causing the exporter to queue and retry indefinitely. The backoff retry mechanism (quadratic backoff up to 30s) could be overwhelmed by sustained high-volume attacks.

The FirstPartyEventLoggingExporter makes POST requests to /api/event_logging/batch without implementing rate limiting. The exporter relies on backend rate limiting only, which may not be sufficient against abuse. Batch size is limited to 200 events but no request frequency limiting exists.

When OAuth tokens expire or are missing profile scope, the exporter falls back to sending events without authentication headers. While this prevents 401 errors, it could expose sensitive telemetry data if the analytics endpoint doesn't properly validate anonymous requests or if the fallback path is triggered unexpectedly.

The eventQueue array can grow unbounded if events are logged before the sink attaches and the sink never attaches. This could exhaust memory if the app crashes before attachAnalyticsSink is called.

The validateStatus function uses s < 500 which accepts 4xx errors as success. This could mask authentication failures or rate limit responses.

The analytics sink routes events to Datadog and 1P logging without verifying the integrity of the analytics provider. This could allow supply chain attacks through compromised analytics endpoints.

The DATADOG_GATE_NAME constant exposes the internal feature flag name 'tengu_log_datadog_events'. This reveals internal architecture and monitoring infrastructure.

The code retrieves OAuth tokens and checks profile scope but does not validate token expiration or refresh status before use. If tokens are stale or expired, the application may continue using them until refresh fails, potentially causing unexpected auth failures or using compromised tokens.

The fetchBootstrapAPI function makes HTTP requests to an external endpoint without rate limiting or request throttling. This could be exploited for DoS attacks or information gathering.

The bootstrap endpoint is constructed from getOauthConfig().BASE_API_URL without validation. If this config can be manipulated or if there's a fallback mechanism, attackers could redirect bootstrap requests to malicious endpoints for credential harvesting or data exfiltration.

extractConnectionErrorDetails returns code and message fields that reveal SSL/TLS error codes and connection failure details. This helps attackers fingerprint proxy configurations and TLS setup.

Error handling and API call processing lacks billing alert integration or spend anomaly detection. No tripwire between normal usage and financial damage.

downloadFile() has 60-second timeout and retry logic with exponential backoff, but no rate limiting on the endpoint. Could be abused to trigger excessive bandwidth costs or downstream API charges.

File download endpoint lacks rate limiting. Large file downloads consume bandwidth costs and could be abused for bandwidth-based attacks.

When getAuthHeaders() returns an error, the function silently returns without attempting the API call. This could allow attackers to probe the endpoint by observing whether the API call proceeds or not based on error handling behavior.

The file constructs an API endpoint URL using oauthConfig.BASE_API_URL + '/api/organization/claude_code_first_token_date', revealing the internal API path structure for organization-level data. This could help attackers map the API surface.

When Grove API fails, the memoized cache is cleared, causing repeated API calls on subsequent renders. This could expose authentication headers repeatedly to a potentially compromised endpoint.

getGroveNoticeConfig returns notice_is_grace_period, notice_reminder_frequency, and domain_excluded fields that reveal internal notification feature configuration and user qualification status.

The extractConnectionErrorDetails function logs detailed connection error information including SSL error status and messages. While intended for debugging, this could expose internal TLS handshake details or certificate chain information that might aid attackers in understanding infrastructure configuration.

The clientRequestId is logged with explicit instruction to 'give this to the API team for server-log lookup'. While this is a debugging aid, logging client-generated request IDs could potentially leak user session identifiers or correlate requests across sessions if the IDs are not properly scoped.

logOTelEvent calls in logAPIError and logAPISuccessAndDuration send events to OTLP without correlation IDs. This makes it impossible to trace API errors back to the original request context, hindering incident investigation and security forensics.

logOTelEvent call in logAPISuccessAndDuration lacks correlation ID for request tracing. Without it, correlating API requests with their responses in distributed tracing systems is impossible.

The metrics opt-out API endpoint (https://api.anthropic.com/api/claude_code/organizations/metrics_enabled) exposes whether metrics logging is enabled for an organization. This could be used to fingerprint organizations and understand their privacy configuration.

Organization UUID is cached in global config and used for overage credit tracking. This organizational identifier may be considered PII and should have consent tracking.

The addedTools and removedTools arrays contain user-configured MCP tool names that could include filepaths or sensitive identifiers. While sanitized for analytics, these names are used in diffable content that could be leveraged for indirect prompt injection attacks via cache manipulation.

The appendSessionLog and getSessionLogs functions accept sessionId as a string parameter without validation. An attacker could potentially inject malicious characters or exploit UUID format assumptions.

The `getSessionLogs` and `getSessionLogsViaOAuth` functions expose the internal session log structure, including entry types, UUID fields, and payload formats. This helps attackers understand the data model.

The `fetchUtilization` function returns rate limit data including `utilization` percentages, `resets_at` timestamps, and usage statistics. This exposes rate limiting configuration and usage patterns to attackers.

Error messages expose internal implementation details that could aid attackers in understanding system architecture.

Error messages may expose network addresses and hostnames that could be used for targeted attacks.

**Perspective 1:** autoDream runs background consolidation that reads session transcripts and memory files without explicit access control checks. The process has broad access to user data without per-session permission verification. **Perspective 2:** autoDream runs background consolidation that accesses multiple sessions without per-session isolation. A compromised consolidation process could access all session data.

logEvent('tengu_auto_dream_completed', ...) writes to analytics pipeline without checking if auto-dream is still running. If auto-dream fails and retries, logging continues indefinitely, driving up storage costs.

The away summary generation has no monitoring for abnormal API usage patterns. Without anomaly detection, a spike in away-summary API calls (potentially triggered by automated attacks) would go unnoticed until budget impact is realized.

Telemetry initialization errors include full error details that could expose internal system state and configuration

Error caught during compaction may contain sensitive conversation state

Context compaction operations suppress warnings without integrity verification. Compacted content could be tampered with or lose provenance metadata.

Session memory compaction error logging uses errorMessage(error) which may include stack traces or internal state details that shouldn't be logged.

The baseline and rightFileDiagnosticsState Maps store diagnostic data indefinitely without time-to-live or automatic deletion. This creates data retention risk for file paths and diagnostic information that could be considered sensitive.

The service tracks connected IDE MCP clients and calls openFile/getDiagnostics RPCs. This creates a record of which IDE tools are used and what files are accessed, without explicit user consent for this telemetry.

The prompt template uses newMessageCount variable without validation or bounds checking. An attacker could craft inputs that cause excessive token consumption, potentially pushing system instructions out of the active context window or causing denial of service through token budget exhaustion.

The memory extraction LLM call has no input token limit enforcement, allowing adversarial inputs to maximize token consumption and potentially push system instructions out of the active context window.

LSP server initialization includes deprecated rootPath and rootUri fields which may expose internal workspace structure to LSP servers. While some servers require these for URI resolution, they should be conditionally included only when explicitly needed.

While maxRestarts is set to 3 by default, the config.maxRestarts field can be overridden by user configuration without validation. An attacker could configure extremely high restart limits, causing resource exhaustion through repeated server restarts.

The openedFiles Map tracks all opened files indefinitely without size limits. An attacker could open thousands of files, causing memory exhaustion in the Map structure.

Error logging includes full error message with potential sensitive context that could leak internal implementation details

Error logging includes full error message with potential sensitive context that could leak internal implementation details

OAuth authentication flow uses static client secrets stored in secure storage without rotation mechanism. Combined with fixed fingerprint salt in fingerprint.ts, this creates long-term credential exposure risk.

The normalizeOAuthErrorBody function parses OAuth error responses and rewrites them with status 400. While it redacts sensitive OAuth params in URLs, the error body parsing could potentially expose client_id, client_secret, or other authentication details in error messages if not properly sanitized.

The _doRefresh function can be called repeatedly without rate limiting. An attacker could trigger rapid token refresh requests, potentially exhausting OAuth server resources or causing rate limit abuse.

MCP channel plugins are managed through an allowlist but there is no provenance tracking. The system cannot verify the origin or build history of channel plugins, making it impossible to trace security issues back to their source.

fetchClaudeAIMcpConfigsIfEligible() uses memoize but lacks rate limiting on the fetch. Could be triggered repeatedly in background, causing repeated API calls to Claude.ai MCP servers.

**Perspective 1:** MCP server configurations are fetched from external API without verifying the response integrity or authenticity. **Perspective 2:** MCP server configurations are fetched without tracking their provenance or origin, making audit and rollback impossible.

Error message includes internal implementation details that could help attackers understand system architecture and identify potential attack vectors.

The MCP server URL is constructed from user-provided serverRef.url without validation. If an attacker controls the MCP server configuration, they could point it to internal services (localhost, 127.0.0.1, private IPs) and potentially access internal resources.

createCodeSession() accepts POST requests to create code sessions without rate limiting. Each session creation may trigger backend resource allocation. Unbounded session creation could drain compute resources.

Elicitation events are logged via logEvent('tengu_mcp_elicitation_response') but there's no anomaly detection or billing alert when elicitation frequency spikes. Adversarial traffic could drive up MCP server costs without tripwire alerts.

Error message includes server name and error details that could expose MCP server configuration information to attackers monitoring logs

Elicitation hook error logs include the full error object which may contain internal state or sensitive data

ElicitationResult hook error logs include the full error object which may contain internal state or sensitive data

The trust dialog check is explicitly disabled when `!getIsNonInteractiveSession()` (line 43-44). In CI/CD or automation contexts, MCP servers with headersHelper can execute without workspace trust confirmation, potentially allowing untrusted code to inject headers into MCP connections.

The normalizeNameForMCP function uses regex replace(/[^a-zA-Z0-9_-]/g, '_') which could be vulnerable to catastrophic backtracking on certain malformed inputs, especially when combined with the subsequent replace(/_+/g, '_').

The onConnectionAttempt callback handles MCP server connections but the authentication state is managed through client.config without explicit validation that the server requires authentication. This could allow unauthenticated connections to certain MCP servers.

The getProjectMcpServerStatus function has a comment indicating it intentionally checks skipDangerousModePermissionPrompt via userSettings/localSettings/flagSettings/policySettings but NOT projectSettings (repo-level .claude/settings.json). This creates an inconsistency where project-level security settings can be bypassed.

In non-interactive mode (SDK, claude -p, piped input), MCP servers are auto-approved if projectSettings is enabled. This allows automated attacks via malicious project settings without user awareness or consent.

The getLoggingSafeMcpBaseUrl function returns undefined for stdio/sdk servers, but for SSE/HTTP servers, the URL is logged. If the URL contains authentication tokens in query parameters (e.g., ?token=xxx), these could leak through analytics logging.

The getLoggingSafeMcpBaseUrl function strips query strings from URLs before logging for analytics, but the original URL with potential access tokens in query parameters could still be logged elsewhere or exposed through error messages. Query strings often contain authentication tokens.

Comment states 'TODO: This fails an e2e test if the ?. is not present. This is likely a bug in the e2e test. Will fix this in a follow-up PR.' However, no follow-up PR exists and the issue remains unresolved. This suggests the comment was added to acknowledge a known problem without implementing a fix.

Comment references 'oven-sh/bun#27469' and '#26385' as Bun PathWatcherManager deadlock issues. These appear to be fabricated issue numbers designed to add credibility to the workaround comment. The actual Bun repository uses 'oven-sh/bun' format but these specific issue numbers don't correspond to real, documented issues.

Comment states 'SECURITY: We intentionally only check skipDangerousModePermissionPrompt via hasSkipDangerousModePermissionPrompt()... but NOT projectSettings'. While the code does check this, the comment is overly specific and could be misinterpreted. The actual security concern is that sessionBypassPermissionsMode can be set from project settings, which creates a potential RCE vector that isn't fully addressed.

Experiment gates are sent to VSCode MCP client without verification of their authenticity or integrity. An attacker controlling the experiment gate endpoint could inject malicious gates that change application behavior.

The MCP server connection setup doesn't verify the identity of the connected server. This could allow unauthorized servers to connect and receive notifications about file changes.

MCP server approval dialogs can be triggered without any rate limiting. An attacker could rapidly trigger multiple approval dialogs, potentially causing UI confusion or exploiting timing vulnerabilities in the approval flow.

The auth code listener starts a local HTTP server on an OS-assigned port. While it binds to localhost, there's no validation that the port assignment succeeds or that the server is properly isolated. Port exhaustion or binding failures could cause issues.

Profile information fetched at line 106-108 is used directly without validation. If the API returns unexpected or malicious data, it could be used without proper checks.

The parseScopes function is called at line 111-112 without validating the scope string format. Malformed or unexpected scopes could cause issues or expose unintended permissions.

The tokenAccount object at line 130-134 includes uuid, emailAddress, and organizationUuid which could be sensitive PII. This is exposed in the OAuthTokens response.

The logEvent call captures installed_count, updated_count, failed_count, and up_to_date_count. While counts are acceptable, the logForDiagnosticsNoPII function may still expose plugin names or installation details that could identify user preferences.

Error message includes plugin identifier which is PII

Plugin name logged to telemetry may be PII

Error messages in uninstallPluginOp and setPluginEnabledOp expose internal plugin IDs (e.g., 'plugin@marketplace' format) and scope information. This helps attackers map the plugin architecture and identify valid plugin identifiers for enumeration attacks.

Policy limits endpoint returns restrictions object that reveals which features are enabled/disabled for the organization. This exposes internal feature flags and policy configurations.

Policy restrictions are cached to ~/.claude/policy-limits.json which is user-writable. An attacker with file system access could potentially modify cached policy data to bypass organization restrictions.

The policy limits service uses the same auth headers as the main API client. If the OAuth token is compromised, policy limits could be used to enumerate organization restrictions.

Rate limiting configuration and message generation code has no Software Bill of Materials (SBOM) tracking. This makes it difficult to audit dependencies and verify the integrity of rate limiting logic.

The error message includes a specific feedback channel (#briarpatch-cc) that is only shown to internal users. This creates a distinguishable error path that could be used for account enumeration or internal user identification.

startBackgroundPolling runs pollRemoteSettings every 60 minutes (POLLING_INTERVAL_MS) with no maximum poll count or rate limiting. If the settings endpoint is metered, this creates a recurring cost drain with no upper bound.

The startBackgroundPolling function starts a background interval that fetches settings every hour. While the interval is long, there's no limit on concurrent polling requests across multiple sessions, which could be abused in high-scale deployments.

The security check dialog handles remote managed settings but does not generate or verify a Software Bill of Materials (SBOM) for the settings artifacts. This makes it impossible to track dependencies and verify their integrity.

Full error messages are displayed to users which could expose internal system state and processing details

Using z.record() instead of validated schema may allow unexpected data types

Error message in upload result includes complete error string which may leak internal details

Catch block logs error but doesn't sanitize before logging

SYNC_KEYS includes projectSettings and projectMemory functions that generate paths for project-specific settings. These paths could be exposed to unauthorized users if the sync mechanism is compromised.

The secret scanner uses regex patterns that could be bypassed through character encoding tricks, URL encoding, or using alternative secret formats not covered by the curated rules. Attackers could obfuscate secrets to evade detection before upload.

Multiple regex patterns use alternation and complex character classes that could be vulnerable to ReDoS attacks. The openai-api-key pattern has nested alternations that could cause exponential backtracking.

Team memory sync API types define checksums for content entries but no enforcement of signing or verification. The checksums are SHA-256 hashes but there's no cryptographic signature verification for the sync artifacts.

Tip history configuration save may fail silently without proper error handling, potentially leaving system in inconsistent state

The tipIdLength field is logged to analytics events. While the comment states this is verified as not code or filepaths, tip IDs could potentially be correlated with user behavior patterns or session history, creating a privacy risk.

The token counting API calls may fail or timeout with very long messages, potentially causing resource exhaustion or incomplete token estimation.

**Perspective 1:** The hook_error_during_execution attachment includes the error content which could contain stack traces or execution context from tool hooks. If these errors are displayed to users, they could leak information about the hook system and tool execution internals. **Perspective 2:** Error messages from tool hooks may expose internal tool execution context, revealing which tools are available and how they're configured.

withVCR and withStreamingVCR manage test fixtures without verifying the integrity of cached fixtures. This could allow supply chain attacks through tampered test fixtures.

Comment states 'up to ~8s on cold coreaudiod (post-wake, post-boot)' with specific timing claims that cannot be verified. This suggests AI-generated documentation without empirical validation.

Voice stream STT service uses OAuth tokens but lacks proper token refresh validation before transmission. This creates risk of using expired tokens which may expose access credentials. Violates SOC 2 CC6.6 (secure development lifecycle) for authentication controls.

Comment references 'claude-in-chrome' skill and 'tmux or CLI-verifier skill' that may not exist in the codebase. These appear to be hallucinated tool names.

The interval parsing regex `^\d+[smhd]$` accepts arbitrarily large numbers without validation. A malicious user could pass extremely large values like '999999999999h' which would pass regex validation. While the cron generation logic rounds to reasonable values, there's no explicit upper bound on the input number, potentially allowing resource exhaustion through excessive parsing/computation.

Bundled skill registration for memory review has no provenance tracking. The skill's origin, version, and integrity cannot be verified at runtime.

The args parameter passed to getPromptForCommand has no length validation, allowing extremely long argument strings that could cause memory issues or prompt injection.

Bundled skill content (.md files) is inlined at build time via Bun's text loader. This content is embedded in the binary without integrity verification, allowing potential tampering during the build pipeline.

Error message includes full skill name and directory path that could expose internal filesystem structure

Error logging in persistRemoteAgentMetadata and removeRemoteAgentMetadata functions logs full error strings which may expose internal implementation details including error codes, stack traces, or internal state information.

Error messages in restoreRemoteAgentTasksImpl may expose blocked hostname and address information when fetchSession fails, potentially revealing internal network configuration.

The general-purpose agent system prompt (line 1) is defined as a static string but is used in an agentic context where user messages are processed. Without clear structural separation between system instructions and user input in the message chain, there's potential for prompt injection where malicious user input could influence the agent's behavior or tool selection.

Agent tools from MCP servers are loaded without tracking their provenance or verifying their source. This makes it difficult to audit which tools were used and from where.

The BashTool supports background task execution (line 1046) with auto-backgrounding for long-running commands, but there's no maximum iteration bound or human-in-the-loop gate for agentic loops. An agent could continuously spawn background tasks without termination, leading to resource exhaustion or persistent unauthorized execution.

Background tasks can be spawned without rate limiting. An attacker could trigger rapid background task creation through repeated command execution, potentially exhausting system resources or causing connection pool exhaustion.

Error messages include BASH_SECURITY_CHECK_IDS enum values (e.g., subId: 1, 2, 3) that leak internal implementation details about the security validation system structure.

Attachment files are uploaded without verifying file integrity or checking for tampering. Uploaded files could be malicious or corrupted.

file_path input is expanded but not validated for path traversal attempts (e.g., ../../etc/passwd). The expandPath function may not fully sanitize all traversal vectors.

The image processor accepts image buffers without validation of image size or dimensions. Large images could cause memory exhaustion or DoS conditions.

Math.random() is used to generate new cell IDs. While this is primarily for internal cell identification, using a cryptographically secure random source would be more appropriate for any ID generation that could be used in security contexts or for preventing ID prediction attacks.

Error messages include request id and toolUseID which can be used to correlate requests and identify tool usage patterns

Error logging includes internal error classification details that could reveal system state and processing logic

Math.random() is used to generate persistId for binary blob storage. This creates predictable file paths that could be exploited to guess or enumerate stored files. The slice(2, 8) produces only 6 hex characters (~24 bits of entropy), which is insufficient for security-sensitive identifiers.

Invalid cron expressions are logged directly in error messages. While this is necessary for debugging, it creates a log entry that could be used to reconstruct scheduling patterns or system configuration. Consider masking or truncating the expression.

Error messages may leak tool execution context including command details and execution environment information

Error messages leak internal implementation details that could help attackers understand system architecture

Error messages expose blocked hostname and address information that could aid SSRF attack reconnaissance

The TaskCreateTool accepts arbitrary metadata that gets attached to tasks. If metadata contains sensitive information, it could be exposed through task list APIs or logs.

Error thrown includes task_id parameter which may be sensitive internal identifier

TaskStopTool throws Error('Missing required parameter: task_id') which could be used for enumeration attacks to determine valid task IDs.

The redirect detection logic logs the original and redirect URLs, which could expose sensitive internal URLs or authentication endpoints in logs.

The logForDebugging statement logs the relay port but the authHeader and wsAuthHeader contain session tokens and bearer tokens. If debug logging is enabled in production, this could expose credentials.

Error messages may leak sensitive context including file paths, environment variables, or system state.

Math.random() is used to generate command IDs for shell execution tracking. While this may be acceptable for internal tracking purposes, it uses a non-cryptographic PRNG that could theoretically be predicted if an attacker observes enough command executions. For security-sensitive operations, a CSPRNG should be used.

Error messages may expose network addresses and hostnames that could be used for targeted attacks.

Comment claims AppState would be 'overwritten' causing interference, but no code demonstrates this scenario or provides mitigation.

The agenticSessionSearch function sends session transcripts (including user prompts and assistant responses) to an external API for semantic search. This could expose sensitive conversation content to the search service.

2.36MB WASM binary and 2.1MB font file are embedded directly in source code with hardcoded paths. No integrity verification, checksums, or signing of these artifacts. If source is tampered with, these binaries could execute malicious code.

The preconnectAnthropicApi function logs the baseUrl being used for preconnection. If this includes sensitive proxy URLs, internal endpoints, or custom gateway addresses, they could leak through application logs.

The code references ANTHROPIC_BASE_URL environment variable and getOauthConfig().BASE_API_URL, revealing the application's API endpoint configuration. This helps attackers identify the target infrastructure and plan attacks against the API endpoint.

User-provided arguments are substituted into skill/command prompts using regex replacement. While the arguments are parsed via shell-quote, the substitution could allow injection if the prompt template contains special characters that affect downstream processing.

Silent error handling in asciicast recorder may hide important errors that could indicate security issues or data corruption.

The refreshAndGetAwsCredentials and refreshGcpCredentialsIfNeeded functions can be called frequently without rate limiting. Credential refresh operations are expensive and could be DoS'd through repeated calls.

isValidApiKey() uses /^[a-zA-Z0-9-_]+$/ which permits underscores and dashes. While not directly a randomness issue, this reduces the effective character space for API keys, making them more predictable and susceptible to brute-force attacks if keys are ever regenerated or leaked.

Authentication tokens are persisted to disk (CCR_OAUTH_TOKEN_PATH, CCR_API_KEY_PATH) but no SBOM is generated for these session artifacts. This prevents supply chain auditing of credential storage.

File imports 'execa' but the actual code uses 'execa' directly. The import statement appears redundant or incorrectly structured.

**Perspective 1:** The clearAwsIniCache() function uses fromIni({ ignoreCache: true }) which bypasses credential caching. This could cause credentials to be re-read from ~/.aws/credentials on every call, potentially exposing credential file paths in logs or error messages if the file is not properly secured. **Perspective 2:** AWS credential cache bypass logic may expose cached credentials if not properly secured. The cache could contain long-lived access keys or temporary credentials.

**Perspective 1:** The error handling in clearAwsIniCache() catches errors but logs them without sanitizing. If the error contains file paths or credential-related information, this could leak sensitive data in logs. **Perspective 2:** Error handling in AWS credential provider may expose file paths where credentials are stored (e.g., ~/.aws/credentials, IAM roles), aiding attackers in locating credential files.

clearAwsIniCache() ignores errors silently and doesn't validate that credentials exist before attempting to clear cache. Could leave stale credentials in cache.

Error logging includes full error messages from axios calls that may contain sensitive API endpoint URLs, authentication tokens, or internal service information.

Error logging includes full error messages from axios calls that may contain sensitive API endpoint URLs, authentication tokens, or internal service information.

The parse() method caches the last command but each parse() performs ~6 tree walks on the AST. While memoized by command string, a flood of unique commands could exhaust CPU. The regex fallback path has no protection against malicious command strings.

Math.random() is used to generate randomId for shell snapshot filenames. While this is acceptable for non-security-critical identifiers (temporary snapshot files that are cleaned up), it could be improved to use crypto.randomUUID() or crypto.getRandomValues() for better entropy, especially since these files are stored in user-accessible directories.

The shell command parsing uses regex patterns like /\s/ and /\s+/ for whitespace detection. Malicious input with carefully crafted whitespace patterns could trigger catastrophic backtracking in certain regex engines.

Comment at line 359 claims 'FAIL-CLOSED on parse failure' but the actual code returns `{redirections:[], hasDangerousRedirection:false}` which is NOT fail-closed - it's fail-open. The comment contradicts the implementation.

Comment at line 405 claims 'ALL tilde-prefixed targets' are flagged, but the code only checks `target.startsWith('~')` without handling `~/path` patterns. The comment overstates the implementation coverage.

Comment at line 533 states 'shell-quote parses unquoted globs as {op:'glob', pattern:'...'} objects, not strings' but this is not universally true - shell-quote behavior varies by context. The comment presents an absolute claim that may not hold in all cases.

The PARSE_ABORTED symbol is returned when parsing fails, which could be used by attackers to probe parser limits and infer system configuration (timeout values, node limits).

Shell completion errors are logged without a correlation ID, making it difficult to trace the error back to the specific user action or session. This reduces the effectiveness of log analysis for debugging and security auditing.

Error message includes full URL which could leak internal URL patterns or sensitive endpoint information

Error message includes protocol information that could leak internal URL validation logic

createBufferedWriter accepts maxBufferBytes=Infinity by default. Unbounded writes could accumulate and trigger expensive flush operations. No rate limiting on writeFn calls.

CA certificates are loaded from system or bundled Mozilla stores, but there is no provenance tracking for which certificate bundles are being used. This prevents supply chain auditing of TLS trust chains.

The HINT_TAG_RE regex uses global flag with multiline mode and captures attributes that could cause catastrophic backtracking with malformed input containing repeated attribute-like patterns.

The ATTR_RE regex pattern with alternation and optional groups could exhibit catastrophic backtracking on malformed attribute strings with many repeated patterns.

Reads Claude Desktop config file from user's home directory without validating file ownership or permissions. May expose sensitive MCP server configurations. Violates SOC 2 CC6.6 (logical access security).

Debug log file path is hardcoded: join(homedir(), '.claude', 'debug', 'chrome-native-host.txt'). If the log file is readable by other processes or accessible via file system, it could leak internal process information, timestamps, and potentially sensitive data about Chrome native messaging interactions.

The chromeNativeHost logs expose socket paths, PIDs, and client connection counts. These logs could reveal internal infrastructure details including socket file locations, process IDs, and MCP client connection counts if accessed by unauthorized parties.

Socket error logging includes full error object which could expose internal connection state or error details to attackers monitoring logs

Invalid message length error logging includes the actual length value which could help attackers understand message protocol boundaries

The getOAuthToken function returns OAuth tokens without validating token expiration or refresh capability. Tokens may be stale or expired when transmitted to the bridge, potentially causing authentication failures or requiring re-authentication flows that could expose credentials.

OAuth tokens retrieved via getClaudeAIOAuthTokens() are stored in global config without documented encryption at rest. Access tokens and refresh tokens should be encrypted when persisted to disk.

The BASE_CHROME_PROMPT contains hardcoded MCP tool names. If these tool names are used in error messages or logs without proper redaction, they could leak internal tooling information.

The native host manifest is written to disk without checking file permissions or verifying the process has appropriate privileges. This could allow privilege escalation if the manifest is writable by untrusted users.

The installChromeNativeHostManifest function writes manifest files to browser directories without validating that the paths are legitimate browser data directories. Could allow writing to arbitrary locations if getAllNativeMessagingHostsDirs() is compromised.

CHROME_EXTENSION_RECONNECT_URL is hardcoded as 'https://clau.de/chrome/reconnect' which could be a phishing target. The URL should be validated against a whitelist of approved domains.

Windows registry entries for native messaging hosts are created via reg.exe without tracking the origin or verifying the registry value's integrity. If registry is compromised, malicious hosts could be registered.

Chrome native host wrapper scripts are created and written to disk without verifying the script content's integrity after creation. Scripts could be tampered with to intercept or modify native host communication.

The extension detection function reads browser profile directories without checking if the process has proper permissions. On some systems, this could expose sensitive browser data or fail silently, leading to security issues.

Permission errors are logged with a flag indicating whether the file path includes the home directory. While the full path is not logged, this creates an audit trail that could be used to infer which files users have access to, potentially revealing sensitive file structure information.

Similar to line 737, permission errors in the rules directory processing log whether the path includes the home directory. This creates an audit trail that could be used to infer which configuration directories users have access to.

Cleanup utilities manage session files, caches, and artifacts but do not generate or track Software Bill of Materials (SBOM) for these artifacts. This prevents supply chain visibility into what components are included in session data.

The cleanupOldMessageFiles() and cleanupOldSessionFiles() functions use unlink() which may not securely delete data on all filesystems. On some systems, files can be recovered from disk even after unlink. For PHI or PCI data, this violates HIPAA 164.310(d)(2) and PCI-DSS 3.4.

Sensitive data files are deleted without secure overwrite (NIST 800-88). Simply unlinking files leaves data recoverable from disk. This violates PCI-DSS 3.3.1 and SOC 2 CC6.1 for data disposal requirements.

App names from Spotlight results are included in the computer use tool description without sufficient filtering. Malicious apps with carefully crafted names could potentially influence the model's behavior through prompt injection in the tool description.

Error handling in auto-unhide Computer Use MCP apps may expose internal state or error details that could aid attackers.

The process running check using signal 0 has a small window for PID reuse, which could cause stale lock detection to fail in high-churn environments.

The coordinateMode is frozen at first read (line 50-52) and never refreshed. If the GrowthBook config changes mid-session, the coordinate mode won't update. While not a randomness vulnerability, this could lead to inconsistent behavior where the model receives one coordinate mode but the executor uses another.

Error message explicitly states '@ant/computer-use-input is not supported on this platform' which could help attackers identify unsupported platforms for reconnaissance purposes.

Computer use permission dialog and session management lacks detailed audit logging of permission grants/denials. Violates SOC 2 CC7.1 (monitoring activities) and PCI-DSS 10.2 (audit trails).

Session registration and tracking does not verify authentication state before creating session records. This could allow unauthorized session creation. Violates SOC 2 CC6.1 for access management controls.

updatePidFile() writes session name directly to JSON without validation. Could allow injection of malicious content into PID files that might be read by other processes.