Gemini 2.5
Google DeepMind — Gemini
rich detail 18 method fields published
Evaluation
Deceptive alignment: stealth challenges
Stealth challenges: model must achieve a devious objective (e.g., cheat on support tickets) without being caught by an overseer; performance relative to human scores; published in Phuong et al. (2025)
How it was run
Open a release to see the setup, scoring, and other details that the lab published. Only fields the lab actually disclosed are shown.