The Agent Skills Directory

[REMOTE_CODE_EXECUTION] (MEDIUM): The skill directs the agent to download and execute code from an untrusted GitHub repository. Although this is the primary mechanism for the tool's operation, the repository owner (martinpacesa) is not a verified trusted organization. Severity is adjusted to MEDIUM per the primary purpose rule. Evidence: git clone https://github.com/martinpacesa/BindCraft.git followed by python bindcraft.py in SKILL.md.\n- [EXTERNAL_DOWNLOADS] (MEDIUM): The skill fetches code and requirements from an unverified external source. Evidence: https://github.com/martinpacesa/BindCraft.git and pip install -r requirements.txt.\n- [COMMAND_EXECUTION] (MEDIUM): The skill executes local Python scripts and cloud-based Modal commands that rely on external content. Evidence: modal run modal_bindcraft.py and python bindcraft.py.\n- [PROMPT_INJECTION] (LOW): The skill ingests structural data from user-provided PDB files without sanitization, creating an attack surface for indirect prompt injection if the agent processes the file contents. 1. Ingestion points: target.pdb in SKILL.md. 2. Boundary markers: Absent. 3. Capability inventory: Local Python and cloud Modal execution. 4. Sanitization: Not documented.

bindcraft