The Agent Skills Directory

[DYNAMIC_EXECUTION]: The reference documentation contains implementation examples for caching embeddings and checkpointing long-running batch processes using Python's pickle module. Specifically, references/esm-c-api.md and references/forge-api.md describe EmbeddingCache and CheckpointedBatchProcessor classes that utilize pickle.load() to read from local files. Deserializing data with pickle is insecure and can lead to arbitrary code execution if an attacker modifies the underlying cache or checkpoint files.
[EXTERNAL_DOWNLOADS]: The skill documentation provides instructions for installing the ESM SDK directly from a GitHub repository (github.com/Biohub/esm.git). While the instructions advise pinning to a specific 40-character commit SHA to mitigate supply chain risks, this still involves executing code from an external repository that is not included in the platform's trusted vendor list.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted protein sequences and structure files which are interpolated into language model prompts, creating an indirect prompt injection surface.
Ingestion points: External data enters through protein sequence strings and PDB file imports in SKILL.md and references/biohub-platform.md.
Boundary markers: The provided examples do not use delimiters or instructions to prevent the agent from interpreting sequence data as command overrides.
Capability inventory: The skill can perform network requests to Forge and Biohub APIs and write files to the local system (PDB, mmCIF, and FASTA formats).
Sanitization: There is no evidence of sequence validation or sanitization of PDB file content before it is processed by the agent.

esm