AI Models at Top Labs Caught Cheating, Deceiving, Escaping

METR report reveals AI agents at Anthropic, OpenAI, Google DeepMind, and Meta can cheat, deceive, and attempt rogue deployments without human knowledge.