alterlab-esm
Warn
Audited by Gen Agent Trust Hub on Apr 12, 2026
Risk Level: MEDIUMREMOTE_CODE_EXECUTIONPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [REMOTE_CODE_EXECUTION]: The documentation and code examples in 'references/esm-c-api.md' and 'references/forge-api.md' suggest using 'pickle.load()' for caching embeddings and checkpointing long-running batch jobs. Pickle is an unsafe deserialization method that can lead to arbitrary code execution if a local cache file (e.g., 'embeddings_cache.pkl') is replaced or modified by an attacker.
- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the ingestion of untrusted protein data and functional annotations. Ingestion points include processing external protein structures via 'ESMProtein.from_pdb' in 'references/esm3-api.md' and 'references/workflows.md'. There are no boundary markers or explicit warnings in the examples to separate external data from agent instructions. The skill possesses capabilities including network requests to 'forge.evolutionaryscale.ai', file system writes for PDB and FASTA files, and package installation. No sanitization or validation logic for the input PDB file content is described in the provided materials.
- [COMMAND_EXECUTION]: Environment setup instructions include installing the 'esm' and 'flash-attn' packages using 'uv pip'. These are standard dependencies required for the skill's core functionality.
- [EXTERNAL_DOWNLOADS]: The skill utilizes 'from_pretrained' methods to download model weights from EvolutionaryScale and Hugging Face repositories. These downloads are directed to well-known official sources and are necessary for local model inference.
Audit Metadata