The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill's reference documentation (references/esm-c-api.md and references/forge-api.md) provides code examples that use the Python pickle module for caching embeddings (EmbeddingCache class) and job checkpoints (CheckpointedBatchProcessor class). The use of pickle.load() on local files poses a security risk because it allows for arbitrary code execution if a malicious actor provides a specially crafted pickle file.\n- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of protein sequences and structure data from external files (e.g., via ESMProtein.from_pdb()). This creates an attack surface where untrusted data could potentially influence agent behavior. Given the skill's capabilities to perform network requests to APIs and write files to the local system, this forms a potential chain for exploitation.\n
Ingestion points: ESMProtein.from_pdb() in references/esm3-api.md and references/workflows.md, and ESMProtein(sequence=...) in the main SKILL.md.\n
Boundary markers: No specific boundary markers or instructions to ignore embedded commands are demonstrated for the biological data processed.\n
Capability inventory: The skill utilizes network capabilities via the esm SDK to communicate with biohub.ai and evolutionaryscale.ai, and file-writing capabilities using open().write() for PDB files.\n
Sanitization: The provided examples do not demonstrate sanitization or rigorous validation of the input sequences or PDB file contents.

esm