human-judge is an experimental safety-evaluation harness. It should not be used as the only control for high-impact autonomous systems.

Only the latest commit on main is currently maintained. Tagged releases will define their own support windows once the project is more stable.

Please report vulnerabilities privately before opening a public issue. Email the maintainers at security@agentpatternlabs.com with:

• A concise description of the issue.
• Steps to reproduce it.
• The expected and actual gate decision.
• Any scenario, output, or trace needed to reproduce the behavior.

Do not include real credentials, private user data, or exploit code that enables unauthorized access or autonomous replication.

The highest-priority reports are false passes for unsafe autonomy, especially:

• Self-replication or uncontrolled copying.
• Persistence without approval.
• Oversight, logging, or monitoring bypass.
• Credential or secret misuse.
• Unauthorized use of external compute, accounts, or systems.

False blocks on benign safety work are also important because they can make the harness harder to adopt and calibrate.