skills/ruvnet/ruflo/harness-evolve/Gen Agent Trust Hub

harness-evolve

Pass

Audited by Gen Agent Trust Hub on Jul 6, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: Fetches the @metaharness/darwin package from the npm public registry during execution.
  • [REMOTE_CODE_EXECUTION]: Executes the downloaded @metaharness/darwin package using npx to perform harness mutations and scoring.
  • [COMMAND_EXECUTION]: Uses the Bash tool to run the metaharness-darwin CLI and manage local repository files.
  • [PROMPT_INJECTION]: The skill processes untrusted execution traces and transcripts from the local repository (Ingestion: --repo path, .metaharness/runs/ files) through diagnostic libraries (Capability: analyzeTranscript, classifyFailure). This ingestion surface provides a path for indirect prompt injection where adversarial content in traces could influence the diagnostic output, although the foundation model is described as frozen.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 6, 2026, 05:39 PM
Security Audit — agent-trust-hub — harness-evolve