OpenAI trained a model to stop refusing hacking questions — and gave it to four security firms first
OpenAI released GPT-5.6-Cyber on Monday, a version of GPT-5.6 Sol trained for vulnerability research and exploit work. On OpenAI's own sensitive-security benchmark it answers 95 percent of the questions the standard model answers 1.5 percent of. The Daybreak programme now splits in two: Blue for defensive work like malware analysis and incident response, Red for offensive research, where the new model lives. OpenAI says it used the model on Chrome's V8 JavaScript engine and found two previously unknown flaws that chain together to break out of V8's sandbox, plus at least five holes in a widely used mobile operating system. Accenture, IBM, CrowdStrike and Cloudflare are the first approved partners; access requires identity checks, monitoring, legal declarations and, from 1 September, a hardware security key.
Related stories
DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best2026-08-21Show the robot once — three to twelve seconds — and it gets the job right 59 times out of 100 with no training at all
2026-08-21The model invents its own tasks, builds the rig to test them, then trains on the results — and DeepReinforce gave the weights away
2026-08-20A free model that fits on one graphics card scores 52 — then spends 22,276 thinking tokens drawing a picture
2026-08-18Zhipu did not build a new model. It kept training the old one — and says coding got 50 percent better.
2026-08-17