harness2
Pass
Audited by Gen Agent Trust Hub on Nov 11, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The helper script
scripts/preflight.pyexecutes thegitbinary usingsubprocess.runto verify the repository root and the current commit hash. This execution is performed with security safeguards, including a restricted environment that whitelists specific system variables and a five-second timeout. This is used solely for repository integrity verification during the mandatory offline preflight check. - [EXTERNAL_DOWNLOADS]: The skill instructions and reference documents describe setup procedures involving downloads from public package registries (PyPI, npm) and official GitHub repositories of well-known organizations. These actions are governed by explicit approval gates and are used to install the agent optimization framework and its research benchmark dependencies (such as LAB and JobBench).
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from external research benchmarks and to refine agent harnesses based on model-generated outputs. This constitutes an ingestion surface for potentially untrusted data; however, the skill provides specific instructions to mitigate risk, such as isolating the execution environment and applying network-layer blocks to GitHub and Hugging Face domains during scored evaluation runs.
Audit Metadata