HarnessAI is in pre-1.0 development. Only the latest main branch receives security updates. Tagged releases (e.g. v0.1.0) are snapshots — upgrade to the latest main or the most recent tag to receive fixes.

Do not open a public GitHub issue for security vulnerabilities.

Send a detailed report to juwon123465@gmail.com with:

1. Affected component — file path or subsystem (e.g. security_hooks.py, install.sh, harness/bin/harness).
2. Impact — what an attacker could achieve (RCE, data leak, privilege escalation, denial of service, etc.).
3. Reproduction steps — minimal command sequence or code snippet.
4. Affected versions — commit SHA or tag.
5. Suggested fix (optional but appreciated).

• Acknowledgement: within 72 hours of report.
• Initial assessment: within 7 days — severity classification (critical / high / medium / low) and disclosure plan.
• Fix: target turnaround depends on severity —
  • Critical (RCE, secret leak): patch within 7 days.
  • High: within 30 days.
  • Medium / low: scheduled for the next minor release.

Reporters are credited in CHANGELOG.md unless they request anonymity.

HarnessAI is a developer-facing tool that runs locally and invokes LLM CLIs (claude, gemini). In-scope issues include:

• Arbitrary code execution via malicious skeleton / profile / agent output
• Secret exfiltration through logs or dashboard
• Shell / SQL / path-traversal injection in subprocess invocation
• Bypass of SecurityHooks or harness integrity gates that results in unauthorized code being committed

Out of scope:

• Vulnerabilities in upstream dependencies (fastapi, uv, Claude CLI, etc.) — report to those projects directly.
• LLM hallucinations or incorrect code output (these are a product concern, not a security vulnerability).
• Issues requiring physical access to the developer's machine.

Thank you for helping keep HarnessAI and its users safe.