The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides code examples for 'EmbeddingCache' and 'CheckpointedBatchProcessor' that utilize the 'pickle' module to load and save data. Evidence: In 'references/esm-c-api.md' and 'references/forge-api.md', the 'pickle.load()' function is used to deserialize cache and checkpoint files. Risk: The 'pickle' module is inherently insecure as it can execute arbitrary code during deserialization. If an attacker provides a malicious '.pkl' file to the agent's environment, loading it via these methods could lead to a full system compromise.
[INDIRECT_PROMPT_INJECTION]: The skill processes complex biological data from external files and user-supplied sequences, creating a surface for indirect prompt injection. Ingestion points: 'ESMProtein.from_pdb()' in 'references/esm3-api.md' and user-provided sequence strings in 'SKILL.md'. Boundary markers: The skill does not implement explicit boundary markers or 'ignore instructions' delimiters for the protein data tracks. Capability inventory: The agent has capabilities to write files ('to_pdb()'), access environment variables ('os.environ'), and execute network requests via the Forge API. Sanitization: No sanitization of protein sequences or metadata is described before processing or inclusion in downstream prompts.
[EXTERNAL_DOWNLOADS]: The skill references installation of the 'esm' SDK from a remote GitHub repository. Evidence: 'references/biohub-platform.md' describes installing the package using 'uv pip install esm@git+https://github.com/Biohub/esm.git@'. Description: The skill follows security best practices by recommending the use of a full 40-character commit SHA to pin the dependency and prevent supply chain attacks.

esm