openfold3-nim
Pass
Audited by Gen Agent Trust Hub on Oct 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes biomolecular sequences and Multiple Sequence Alignments (MSAs) which are interpolated into JSON payloads for structure prediction requests.
- Ingestion points: Data enters through the
sequence,smiles,ccd_codes, andmsa.alignmentfields inSKILL.mdandreferences/api.md. - Boundary markers: Instructions do not mandate the use of delimiters or escaping when the agent incorporates user data into the request body.
- Capability inventory: The skill performs network operations via
requests.postand local file writes usingopen()inSKILL.mdandreferences/examples.md. - Sanitization: No explicit sanitization of sequence or alignment strings is performed before payload construction.
- [COMMAND_EXECUTION]: The skill provides shell commands for local container orchestration and service monitoring.
- Evidence: Includes scripts in
SKILL.mdfor registry authentication, launching the NIM container withdocker run, and executingcurlloops for readiness checks. - [EXTERNAL_DOWNLOADS]: Fetches required deployment artifacts from vendor-controlled infrastructure.
- Evidence: Downloads the
openfold3NIM container image and model weights (10–15 GB) fromnvcr.ioas documented inSKILL.md.
Audit Metadata