The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses pickle.load() to deserialize local data in multiple instances:
In references/esm-c-api.md, the EmbeddingCache class loads data from embeddings_cache.pkl.
In references/forge-api.md, the CheckpointedBatchProcessor and ForgeCache classes use pickle.load() for checkpoint.pkl and cache files in forge_cache/.
While used for local caching, pickle is susceptible to arbitrary code execution if the data files are modified by a malicious actor.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill instructs users to install external packages from official registries:
esm (legitimate package for this functionality).
flash-attn (standard library for optimized attention mechanism).
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted protein sequence data and structural files (PDB) which could potentially contain malicious patterns designed to influence the model's output or the agent's behavior.
Ingestion points: ESMProtein constructor and from_pdb method in references/esm3-api.md.
Boundary markers: No specific delimiters or warnings for sequence data ingestion are provided.
Capability inventory: Subprocess calls (via package installation), network operations (Forge API), and file writing (PDB and sequence files).
Sanitization: No explicit sanitization or validation of the input sequences or PDB files is shown before processing.

esm