The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads a Docker container from the vendor's official registry (nvcr.io/nim/mit/boltz2:1.6.0). This is a legitimate operation for local NIM deployment from a recognized service.
[COMMAND_EXECUTION]: Provides instructions to execute shell commands for Docker management and local directory setup. The use of chmod 755 on a locally created cache directory is a standard permission setting for application data and does not constitute a privilege escalation risk.
[CREDENTIALS_UNSAFE]: Instructions correctly direct the agent to read secrets from environment variables (NGC_API_KEY) and explicitly advise against exposing these credentials in logs or terminal output. Sourcing a .env file for local development is a standard and safe practice for secret management.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied protein sequences and chemical strings (SMILES). 1. Ingestion points: sequence and smiles fields in polymers/ligands (SKILL.md). 2. Boundary markers: relies on JSON serialization for API requests. 3. Capability inventory: Bash (Docker), Write (CIF files), Read (Local files). 4. Sanitization: Not explicitly implemented in instructions beyond structured JSON formatting. This surface is characteristic of the skill's scientific purpose.

boltz2-nim