grey-haven-evaluation

Pass

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [Indirect Prompt Injection] (LOW): The skill implements an evaluation pattern that ingests untrusted output from other LLMs.
  • Ingestion points: Untrusted data is processed via the {output} variable in prompts defined in reference/llm-as-judge-guide.md.
  • Boundary markers: The templates use Markdown headers to delimit untrusted content from judge instructions.
  • Capability inventory: The skill is authorized to use the Bash and Write tools according to the SKILL.md frontmatter.
  • Sanitization: Implementation guides suggest using JSON schema validation with pydantic or zod to ensure the integrity of the judge model output.
  • [COMMAND_EXECUTION] (SAFE): The skill permits the use of the Bash tool for running local evaluation scripts, which is appropriate for a testing and performance benchmarking toolset.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 17, 2026, 06:39 PM