rl-env-from-description
Pass
Audited by Gen Agent Trust Hub on Sep 1, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill fetches configuration and dependencies from well-known and trusted sources, including NVIDIA's Gym repository (github.com/NVIDIA-NeMo/Gym), Hugging Face's OpenEnv (github.com/huggingface/OpenEnv), and PrimeIntellect's Verifiers (github.com/PrimeIntellect-ai/verifiers).
- [COMMAND_EXECUTION]: The skill executes generated environment code and rollout scripts via subprocesses to perform 'smoke tests'. This automated validation ensures the generated code is runnable but involves executing user-defined domain logic.
- [INDIRECT_PROMPT_INJECTION]: The skill features a surface for indirect prompt injection as it processes untrusted user data to generate and subsequently execute code.
- Ingestion points: Plain-English environment descriptions and reward criteria are ingested via the main prompt and interview bank (SKILL.md, references/interview.md).
- Boundary markers: No explicit delimiters or isolation instructions are provided to separate user-supplied descriptions from the agent's internal code generation logic.
- Capability inventory: The skill maintains the capability to write multiple files (domain logic, framework implementations) to the local filesystem and execute them via shell commands for rollout testing.
- Sanitization: The skill does not describe any specific validation or sanitization steps for user-provided descriptions before they are used to build executable code.
Audit Metadata