openfold3
Pass
Audited by Gen Agent Trust Hub on Sep 22, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads model weights from the OpenFold3 repository on Hugging Face using standard CLI tools. It also installs Python dependencies from the official PyPI registry.
- [DATA_EXFILTRATION]: By default, biological sequences are sent to the ColabFold MSA server (api.colabfold.com) and the RCSB PDB database (data.rcsb.org) to perform Multiple Sequence Alignment and template searching. The documentation explicitly notes this behavior and provides an opt-out flag.
- [COMMAND_EXECUTION]: The skill relies on several CLI tools to function, including pip for dependency management, huggingface-cli for weight retrieval, and apt-get for installing necessary system libraries like libxrender1.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data via queries.json.
- Ingestion points: Biological sequences and chemical structures are read from queries.json.
- Boundary markers: Input is strictly validated using Pydantic (InferenceQuerySet) with the extra: forbid configuration to prevent unexpected fields.
- Capability inventory: The skill executes the run_openfold prediction engine, weight management via huggingface-cli, and system package management via pip and apt-get.
- Sanitization: Validation is handled by the InferenceQuerySet schema, ensuring data conforms to expected biological sequence formats before processing.
Audit Metadata