Tag: METR

Could a single prompt make AI dangerous? What safety researchers have actually found

The debate between Yoshua Bengio and Yann LeCun on AI risk is now informed by real evaluations of jailbreaks, deceptive behavior and agents that act in the world.

Jul 17, 2026

Models & Research

Safety testers say GPT-5 shows both rapid capability gains and persistent deception

Independent evaluators at METR, Apollo Research and Haize Labs describe what frontier models can now do, and why they still worry, in a New York Times report by Stephen Witt.

Oct 12, 2025