physicalai-train-benchmarking-a-policy

Pass

Audited by Gen Agent Trust Hub on Sep 6, 2026

Risk Level: SAFE
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill facilitates the dynamic loading of policy classes through CLI arguments (e.g., --policy physicalai.policies.ACT). This is a standard design pattern for flexible model evaluation frameworks and relies on the vendor's internal package structure.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources to define its behavior.
  • Ingestion points: Model checkpoints (.ckpt files) and benchmark configuration files (.yaml) are loaded at runtime.
  • Boundary markers: No explicit delimiters or instructions are provided to the agent to ignore embedded commands within these files.
  • Capability inventory: The skill executes policy rollouts in simulation environments, writes results to file system locations (results.json, results.csv), and records qualitative videos.
  • Sanitization: The instructions do not detail specific validation steps for the binary checkpoint data or the structure of the YAML configurations beyond standard loading practices.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 6, 2026, 01:43 PM
Security Audit — agent-trust-hub — physicalai-train-benchmarking-a-policy