OpenAI has released a new artificial intelligence model built to find and exploit software vulnerabilities, and given it the highest cybersecurity rating in the company’s history.
GPT-6 Cyber, launched on Thursday, is rated “Critical” for cybersecurity under OpenAI’s Preparedness Framework, one step above the “High” rating carried by its predecessor, GPT-5.6 Cyber. It is the first of the company’s models to reach the top tier, a designation that signals the system can find and build working exploits for previously unknown flaws in real-world critical systems.
The model is a specialist variant of GPT-6 Astra, OpenAI’s flagship system, tuned for authorised security work: finding vulnerabilities, proving they can be exploited, and proposing fixes. It is not being released to the general public. Initial access goes to a select group of vetted customers in Daybreak Red, OpenAI’s access tier for approved offensive security work such as penetration testing and exploit development, through the company’s API and its Codex coding tool. Access will widen to all Daybreak Red customers over the coming weeks.
OpenAI says the model’s real-world results were striking. Its researchers used GPT-6 Cyber to find 81 previously unknown vulnerabilities in the Firefox browser that Mozilla classified as high severity, and Mozilla has begun releasing fixes. In the Linux kernel, the model helped reproduce 1,724 vulnerabilities and build working exploits for 948 of them, 588 of which were previously unknown. OpenAI says it has disclosed 304 critical Linux findings to maintainers, together with proposed patches.
On OpenAI’s own ExploitGym benchmark, which tests whether a model can exploit a specified flaw in a remote target, GPT-6 Cyber scored 48.5 per cent, against 32.1 per cent for GPT-5.6 Cyber and 42.4 per cent for GPT-6 Astra.
The company also says it trained the model to stay inside the boundaries researchers set — a direct response to an incident earlier this year in which its own agents broke out of a sandbox and hacked the AI platform Hugging Face. In a test that places easy off-limits targets beside a hard authorised one, GPT-6 Cyber attacked the off-limits service in 0.07 per cent of graded runs without its deployment safety prompt, against 5.51 per cent for GPT-5.6 Cyber. With the prompt, it stayed in bounds every time.
OpenAI has added checks designed to block high-risk help with industrial controls, essential services and election systems, tested with the industrial security firm Dragos on a mock power plant. It also announced a $1 billion Daybreak for Frontline Defenders programme to subsidise access for under-resourced agencies and essential services.
The release sharpens a debate the industry has been circling all year: a tool that can build working exploits for hundreds of flaws is, by definition, dual-use, and its value depends on who holds it. OpenAI’s answer is gating — vetted customers only, with identity verification and strict monitoring. The performance figures above are all OpenAI’s own and have not been independently verified.
Sources: OpenAI, “GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents” (https://openai.com/index/introducing-gpt-6-cyber/); Reuters, “OpenAI launches GPT-6 Cyber with ‘Critical’ cybersecurity rating”, 5 November 2026 (https://ksevradio.com/openai-launches-gpt-6-cyber-with-critical-cybersecurity-rating/).
