skill-eval-improve

Pass

Audited by Gen Agent Trust Hub on Jul 30, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill facilitates an evaluation loop that processes external data, including user-provided prompts and third-party skill files. It lacks guidance on utilizing boundary markers or sanitization techniques when the agent interpolates this untrusted content into its instructional context.
  • Ingestion points: Ingests external SKILL.md content and user prompts defined in evals/cases/*.yaml and references/evals.md for evaluation.
  • Boundary markers: Absent; the instructions do not specify the use of delimiters (e.g., XML tags) to separate untrusted data from agent instructions.
  • Capability inventory: The skill utilizes shell commands for validation (pnpm run validate), static analysis (plugin-eval), and benchmarking (steward benchmark).
  • Sanitization: Absent; no filtering or escaping of ingested data is prescribed.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 30, 2026, 10:35 AM
Security Audit — agent-trust-hub — skill-eval-improve