fine-tuning-serving-openpi

Pass

Audited by Gen Agent Trust Hub on Sep 9, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: Fetches the OpenPI repository from GitHub and model checkpoints from Google Cloud Storage assets.
  • [COMMAND_EXECUTION]: Executes training, serving, and utility scripts using the uv environment manager.
  • [REMOTE_CODE_EXECUTION]: Performs installation of the openpi-client package directly from a remote Git URL.
  • [DYNAMIC_EXECUTION]: Directs users to manually overwrite source files in the installed transformers package, which modifies the library's runtime behavior.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external datasets and model weights. Ingestion points: External dataset repositories (e.g., repo_id) and remote checkpoints. Boundary markers: None identified. Capability inventory: Shell script execution and network-based policy serving. Sanitization: None identified.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 9, 2026, 07:08 PM
Security Audit — agent-trust-hub — fine-tuning-serving-openpi