OpenAI released GPT-6 Astra with 100% on ExploitBench, prompting the lab’s first Critical cyber designation, a limited rollout and White House vetting. Trained on more than 100,000 GPUs, it had Greg Brockman declaring: welcome to the AGI era.
The catch is legibility. A reported recurrent-depth trick hides more of its thinking, and one researcher’s view is that legible reasoning was always doomed.