llm-regression-runner
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data from golden datasets and model outputs, which are then used as context for LLM-as-judge prompts, creating an attack surface for indirect instructions to influence the agent.
- Ingestion points: The agent is instructed in Step 1 to search for and read golden datasets (JSON, JSONL, CSV), judge prompts, and context files.
- Boundary markers: No specific delimiters or "ignore instructions" wrappers are prescribed for the interpolation of dataset entries or model outputs into the judge's context.
- Capability inventory: The skill involves reading project files, writing to the file system (Step 6), and executing evaluation logic or programmatic scripts (Step 3).
- Sanitization: There are no instructions to sanitize or validate the content of the external data before it is processed by the agent or the judge prompts.
- [DYNAMIC_EXECUTION]: The skill identifies and executes local evaluation scripts and programmatic rules based on project content.
- Evidence: Step 3 directs the agent to "run the check against the output" for programmatic rules and to execute existing "eval scripts" found during the artifact search in Step 1.
Audit Metadata