The Agent Skills Directory

[COMMAND_EXECUTION]: The reference documentation provided in the skill recommends using the 'pickle' library for local data persistence and caching, which is an unsafe practice.
Evidence: The 'EmbeddingCache' class in 'references/esm-c-api.md' and the 'CheckpointedBatchProcessor' in 'references/forge-api.md' both utilize 'pickle.load()' to deserialize local files ('embeddings_cache.pkl' and 'checkpoint.pkl').
Risk: The 'pickle' module is not secure against erroneous or maliciously constructed data. Never unpickle data received from an untrusted source, as it can be exploited to execute arbitrary shell commands during the deserialization process.
[EXTERNAL_DOWNLOADS]: The skill relies on external software packages and remote service connectivity for its primary functions.
Evidence: The installation instructions in 'SKILL.md' require 'uv pip install esm' and 'flash-attn'.
Evidence: The 'ESM3ForgeInferenceClient' established in 'SKILL.md' and 'references/forge-api.md' connects to 'https://forge.evolutionaryscale.ai' for cloud-based model inference.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by processing complex biological data formats that can be influenced by external sources.
Ingestion points: The skill ingests untrusted data via 'ESMProtein.from_pdb()' (reading PDB files) and protein sequences provided as strings in 'SKILL.md' and 'references/workflows.md'.
Boundary markers: No explicit boundary markers or 'ignore embedded instructions' warnings are present to delineate data from processing logic.
Capability inventory: The skill has the capability to write files (e.g., 'novel_gfp.pdb' in 'references/workflows.md'), perform network operations (Forge API), and execute complex Python logic.
Sanitization: There is no evidence of sanitization or validation for the incoming protein sequences or PDB structures before they are processed by the underlying models.

esm