The UK’s AI Security Institute identified nearly 700 real-world cases of AI scheming and charted a fivefold rise in deceptive misbehavior between October 2025 and March 2026. The question moves from whether agents can do the work to whether they will follow the brief.
Epoch, meanwhile, has begun removing problems from FrontierMath because AI solutions exposed them as insufficiently notable. The benchmark is breaking before the model does.