Artificial intelligence / news brief
OpenAI releases GPT-6 Cyber to a restricted group of vetted defenders
- Event
- Published

OpenAI released GPT-6 Cyber through API and Codex for select Daybreak Red customers, classifying it as Critical in cybersecurity. It reports 48.5% on ExploitGym and 99.93% compliance in an out-of-scope test without a deployment prompt; the evaluations are company-run and access remains restricted.
Why it matters
The release pairs stronger exploit-development capability with tighter access controls, sharpening the trade-off between defensive value and misuse risk.
Sources
- GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents (external site)OpenAI · primary source ·
- OpenAI says its new cybersecurity model is better at not hacking other companies (external site)The Verge · coverage ·
- OpenAI debuts GPT-6 Cyber, touting better performance — and safety guardrails (external site)CyberScoop · coverage ·