The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill directs the cloning and installation of software from an untrusted GitHub repository (github.com/huifer/skill-security-scan). The discrepancy between the repo owner (huifer) and the identified author (reason-machines) is a significant security red flag for supply-chain integrity.
[REMOTE_CODE_EXECUTION]: The skill instructions facilitate the execution of the downloaded third-party tool through shell commands and Python subprocess examples. Since the tool originates from an untrusted and unverifiable source, it provides a pathway for the execution of potentially malicious code.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external, untrusted data (other Claude Skills). This creates an attack surface where malicious content within the analyzed files could influence the agent's internal logic or output.
Ingestion points: File system paths and directories provided to the scan command for analysis (e.g., SKILL.md and associated scripts).
Boundary markers: None identified; the skill processes raw file content for pattern matching without explicit isolation.
Capability inventory: The skill utilizes subprocess.run, file system access, and network requests, while also generating multi-format reports (HTML/JSON).
Sanitization: No specific sanitization, escaping, or filtering mechanisms for the external data being analyzed are mentioned in the skill description.
[COMMAND_EXECUTION]: The documentation provides explicit commands for shell execution, including git clone, pip install, and pyinstaller. These commands are used to acquire and build the external tool, leading to the execution of code from an untrusted source.

skill-security-scanner