The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill installs software packages directly from external GitHub repositories (github.com/Biohub/transformers and github.com/Biohub/esm) rather than exclusively from the official PyPI registry. This is a common practice for research models to include specific patches or features required for specialized performance.
[REMOTE_CODE_EXECUTION]: The skill implementation includes a monkeypatch for torch.linalg.svd to ensure numerical stability when running on GPU hardware. This involves dynamically overriding a standard library function at runtime to redirect specific calculations to the CPU when degenerate inputs are detected.
[DATA_EXFILTRATION]: The skill requires network egress to communicate with HuggingFace Hub for downloading model weights and accesses the Biohub API (biohub.ai) to retrieve feature metadata for ESMC models.
[PROMPT_INJECTION]: The skill processes untrusted biological sequence data (Protein, DNA, RNA, SMILES) which enters the agent context via StructurePredictionInput, presenting a surface for indirect prompt injection attacks.
Ingestion points: The StructurePredictionInput sequence fields and MSA file loading (MSA.from_a3m) in SKILL.md serve as points where external data enters the environment.
Boundary markers: No explicit boundary markers or instructions to ignore embedded commands were identified for the sequence inputs.
Capability inventory: The skill has the capability to write local files (e.g., best.complex.to_mmcif()) and perform network operations to HuggingFace and the Biohub API.
Sanitization: No specific sanitization or validation of the biological sequence content (beyond standard model processing) was identified.

esmfold2