harness-retro

Pass

Audited by Gen Agent Trust Hub on Oct 26, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data sources to generate its recommendations, creating an attack surface for indirect prompt injection.
  • Ingestion points: The skill reads the current conversation history, sub-agent reports, and past session logs or transcripts (SKILL.md, Step 1).
  • Boundary markers: No explicit boundary markers or instructions to disregard embedded instructions are specified for the analyzed data.
  • Capability inventory: The skill can modify project configuration files (AGENTS.md, rule files, scripts) and run repository commands (SKILL.md, Step 5).
  • Sanitization: The instructions do not describe sanitization or validation of the content ingested from logs or reports.
  • Context: The risk is significantly reduced by a human-in-the-loop requirement where the user must approve findings before any changes are applied.
  • [COMMAND_EXECUTION]: The skill is instructed to execute the repository's own verification commands (such as linters or test suites) to ensure that newly proposed harness improvements work as intended (SKILL.md, Step 5).
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 26, 2026, 10:29 AM