T01 · Skill Instruction Hijacking
- Location
discover.mjs:43- Finding
Untrusted CLAUDE.md Content Is Loaded for Privileged Instruction Injection
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is not overtly destructive, but it automatically reads local CLAUDE.md files and promotes their contents into privileged agent context without clear user control or containment.
Install only if you intentionally want CLAUDE.md files from the current directory and nearby parent directories to influence the agent. Review those files and any symlinks first, and prefer an opt-in version that confines reads to an approved project root and treats loaded content as untrusted reference material rather than system instructions.
discover.mjs:43Untrusted CLAUDE.md Content Is Loaded for Privileged Instruction Injection
discover.mjs:17Discovery Escapes the Project Boundary and Can Follow External Filesystem Links
The skill is configured to auto-execute on every user message without manual invocation, causing automatic file discovery and context injection even when unrelated to the user's request. In this context, that makes the behavior more dangerous because it creates persistent, non-consensual exposure of local project content and increases the chance of prompt injection from repository-controlled files influencing all subsequent interactions.
The skill explicitly states it will traverse directories, read CLAUDE.md files, process @include directives, and inject the resulting contents into system context, but it provides no user-facing warning or consent mechanism. This is dangerous because project files may contain sensitive instructions or data, and silently elevating file contents into system-level context can change model behavior in ways the user did not intend.
The script automatically discovers local CLAUDE.md files, reads their contents, and prints them specifically for injection into system context with no trust boundary checks, provenance validation, or user warning. Because these files can contain arbitrary natural-language instructions, this creates a prompt-injection supply-chain risk where repository-controlled content can silently influence downstream agent behavior.
The script automatically discovers and prints the full contents of CLAUDE.md files from the current and parent directories to stdout, with no confirmation, redaction, or warning. If those files contain secrets, internal instructions, or sensitive project context, invoking this utility can unintentionally disclose that data to logs, calling processes, or users who did not expect file contents to be emitted.
The file's user-facing natural-language description is entirely in Chinese, with no indication that other languages are supported or that Chinese is a required locale for a region-specific tool. Under the policy, forcing a specific language without user opt-in or justification is a locale-policy issue.
The human-readable comments and usage guidance are written in Chinese only, with no indication that language choice is optional or contextually required. Under the stated policy, forcing a specific language without user opt-in can be a locale-policy violation.