skill-eval
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill is designed to execute a local Python script
skills/skill-eval/scripts/eval_runner.py, which manages the lifecycle of agent evaluations by spawning subprocesses for executors and graders. This behavior is the primary intended purpose of the skill and is subject to explicit user authorization before execution. - [INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it processes untrusted output from AI models during the grading phase.
- Ingestion points: Evaluation results and executor outputs (captured from external LLM providers) are ingested into the grader's context.
- Boundary markers: The runner employs 'inert sentinels' and explicit instructions to the grader to treat executor output as untrusted data and ignore any embedded instructions.
- Capability inventory: The runner can execute subprocesses (
subprocess.run), manage file system artifacts, and interact with network-based LLM providers. - Sanitization: The system uses JSON escaping for artifact rendering and enforces strict schema constraints on grader outputs to prevent malformed data from affecting the orchestration logic.
Audit Metadata