The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in SKILL.md direct the agent to detect the host operating system and use package managers such as apt, brew, pip, and npm to install missing dependencies. It explicitly recommends the use of sudo for Linux distributions (apt/dnf) to perform these installations.\n- [REMOTE_CODE_EXECUTION]: The scripts/tools/runner.mjs script utilizes execFile to run external audit tools. While the skill provides a registry of known tools in scripts/tools/registry.mjs, the instructions in SKILL.md empower the agent to select and install 'any tool that would be particularly useful' for the detected ecosystem, creating a vector for the execution of untrusted software.\n- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection (Category 8). \n
Ingestion points: The scripts/utils/fs-walk.mjs and scripts/utils/line-reader.mjs modules recursively read content from the user's codebase.\n
Boundary markers: Absent. No specific delimiters or instructions are used to separate raw code content from agent instructions in the reporting phase.\n
Capability inventory: The tool has file-write capabilities via scripts/audit.mjs (writing to CODE_AUDIT_REPORT.md) and shell execution via scripts/tools/runner.mjs.\n
Sanitization: Absent. Extracted code snippets from the audited files are included directly in the markdown report without sanitization, which could lead to malicious content being presented as authoritative audit findings.

code-audit