openfold3-nim

Pass

Audited by Gen Agent Trust Hub on Aug 13, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: Provides shell commands to automate the deployment of the OpenFold3 NIM via Docker and verify service availability using health check endpoints.
  • [EXTERNAL_DOWNLOADS]: Fetches the OpenFold3 NIM container image and required model weights from the official NVIDIA Container Registry (nvcr.io). These are recognized as trusted vendor resources.
  • [PRIVILEGE_ESCALATION]: The local setup script utilizes chmod 777 on the LOCAL_NIM_CACHE directory. While intended to resolve permission conflicts between the host and the container, this is an insecure practice that grants global read, write, and execute permissions to the directory.
  • [INDIRECT_PROMPT_INJECTION]:
  • Ingestion points: User-provided biological sequences (protein, DNA, RNA) and SMILES strings are ingested and processed in SKILL.md and references/examples.md.
  • Boundary markers: The skill uses standard FASTA headers (e.g., >query\n) to delimit sequence data within the request payload.
  • Capability inventory: The skill possesses Bash execution capabilities for Docker management and uses requests.post for outbound network communication.
  • Sanitization: The skill lacks explicit sanitization or structural validation for sequences before they are interpolated into the API request alignment fields, creating a potential surface for injection via malformed sequence data.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 13, 2026, 07:52 AM
Security Audit — agent-trust-hub — openfold3-nim