Security fixes are accepted for the current v1.3.x release line.

Please use GitHub's private vulnerability reporting flow from the repository's Security tab. Do not open a public issue for:

• credentials or private data found in tracked files or artifacts;
• a way for an evaluated agent to read graders, private seeds, verifier tokens, or mock source;
• a mock-service write or verifier-state bypass;
• container escape, privilege escalation, or cross-run state leakage; or
• result tampering that can produce an invalid passing score.

Include the affected commit, task ID, runtime digest, minimal reproduction, and impact. Do not include real production credentials or customer data.

Treat every evaluated model as untrusted code. The agent has shell access in its task container, and provider credentials must be present in that runtime long enough to make model calls. Use dedicated, least-privilege evaluation keys with spend and rate limits. Do not grant them access to production resources, and rotate or revoke them after use.

Benchmark-quality bugs that do not expose a security or integrity boundary can be filed as ordinary issues.