policy-evaluation

Pass

Audited by Gen Agent Trust Hub on Nov 18, 2026

Risk Level: SAFE
Full Analysis
  • [SAFE]: The skill outlines a formal evaluation framework focusing on policy correctness, quality signals, and safety invariants. It promotes security-centric methodologies such as 'Cold Evaluation' (testing in fresh, isolated contexts) to prevent prompt leakage and bias.
  • [COMMAND_EXECUTION]: The skill references the execution of local scripts for deterministic verification, such as bun evals/policy-evaluation/context-budget.mts and shell scripts like grade.sh. These operations are performed on repository-local files for environment-specific checks (e.g., context size monitoring) and do not involve remote code execution or unauthorized privilege escalation.
  • [INDIRECT_PROMPT_INJECTION]: While the skill is designed to process untrusted data (artifacts and evaluation evidence), it specifically mitigates injection risks by enforcing boundary markers, 'fresh context' isolation, and 'blind judging.' These controls ensure that instructions embedded in data do not influence the agent's core system prompts or evaluation logic.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 18, 2026, 12:11 AM