Skip to content

Security: TIGER-AI-Lab/RewardHarness

Security

SECURITY.md

Security policy

Reporting a vulnerability

If you find a security issue (leaked credentials in code or commit history, command injection, etc.), please do NOT open a public GitHub issue.

Instead, email the maintainers at reacher.z@pm.me with:

  • A short description of the issue
  • Reproduction steps or proof-of-concept (commit SHA, file path, exploit input)
  • Whether the issue is currently public anywhere
  • Your suggested fix, if any

We aim to acknowledge reports within 5 business days and ship a patch (or a documented mitigation) within 14 business days.

Supported versions

Only the current main branch and the most recent tagged release receive security fixes. We will not backport fixes to older tags.

Version Supported
main
v0.1.2 (latest)
earlier

Disclosure history

Date Issue Fixed in
2026-05-16 A hardcoded internal Gemini-gateway API key was inadvertently shipped in vanilla/bench_wanqing.py and three vanilla/gemini_bench_*.py files in the initial v0.1.0 release. The key was rotated upstream; current vanilla/gemini_bench_*.py reads credentials from GEMINI_GATEWAY_API_KEY instead. The leaked value remains in git history before commit 6b61e3d — treat any token captured during that ~24 h window as compromised. v0.1.1 (6b61e3d)

Hardening notes

  • .gitignore excludes every .env* except .env.example. Don't commit real .env files.
  • The test suite is fully mocked and must never make real network calls. If a test reaches out, file a bug — it's a security smell.
  • scripts/check_env.py validates that credentials look well-formed but never logs them.
  • Service-account JSON files belong outside the repo tree. Reference them via GOOGLE_APPLICATION_CREDENTIALS only.

There aren't any published security advisories