generate-nemo-gym-env
Pass
Audited by Gen Agent Trust Hub on Oct 7, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill facilitates the installation of the
nemo_gymlibrary from NVIDIA's official GitHub repository.- [INDIRECT_PROMPT_INJECTION]: The generated environment server includes logic to process model-generated output for reward calculation, which is a standard requirement for reinforcement learning environments but represents an indirect injection surface. - Ingestion points: Data enters the system via the
/verifyendpoint inserver.py, specifically through thebody.response.outputfield which contains model trajectories. - Boundary markers: The generated template does not implement specific delimiters or instructions for the agent to ignore potentially malicious instructions embedded in the trajectories.
- Capability inventory: The server manages session states and executes tool-specific logic based on model interactions through the
SimpleResourcesServerclass. - Sanitization: The provided reward logic uses substring matching or type checks on raw model output without additional filtering or sanitization.
Audit Metadata