Google

2026 — Present

Tracks cases in which frontier AI agents exceed intended evaluation or test boundaries, access real systems, or exhibit control failures or unintended behavior, along with disclosures, mitigations, and improvements to evaluation and safety practices.