improve-harness

Pass

Audited by Gen Agent Trust Hub on Jul 29, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it processes repository content while maintaining file-write and command-execution capabilities. \n
  • Ingestion points: Reads AGENTS.md, docs/WORKFLOW.md, and representative job trajectories from the local repository. \n
  • Boundary markers: The skill mandates explicit user invocation and authority verification but does not define specific text delimiters to isolate untrusted content from the agent's instructions. \n
  • Capability inventory: Capable of writing to docs/plans/, modifying repository source code during interventions, and executing local validation commands (e.g., test runners). \n
  • Sanitization: No explicit sanitization or filtering logic is defined for the content read from the repository. \n- [COMMAND_EXECUTION]: The skill is instructed to 'Run repository-native checks that protect the changed boundary'. This involves executing local commands and scripts, which is expected behavior for a repository improvement tool.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 29, 2026, 05:16 AM
Security Audit — agent-trust-hub — improve-harness