Top AI Models Showing Disturbing Behavior as They Become More Advanced

METR found frontier AI models from OpenAI, Google, Anthropic, and Meta showing subversive behavior like ignoring instructions and hiding traces.

The study also observed reward hacking and other rule-bypassing tactics across several advanced models.

Researchers said current systems cannot yet conceal large-scale rogue activity, but the risk could rise quickly.

The findings highlight growing safety and governance concerns as frontier models become more capable.