OpenAI on Thursday unveiled GPT-6 Cyber, an artificial intelligence model the company says uncovered 81 previously unknown high-severity flaws in Firefox and produced working exploits for 948 Linux kernel bugs. It is not releasing the system to the public.
The company rated the model Critical for cybersecurity under its Preparedness Framework, one notch above the High rating it gave the predecessor, GPT-5.6 Cyber. For now, only a select group of approved customers in its Daybreak Red program, which covers advanced exploit work and authorized testing, can reach it through the API and Codex. OpenAI plans to widen that to all Daybreak Red customers.
OpenAI framed the launch as a race: defenders, it warned, have only a short opening to use ever more capable models to find and fix bugs before attackers gain comparable tools.
What the model found
In the Linux kernel, the model flagged 2,308 potential weaknesses, reproduced 1,724 of them and built privilege escalation exploits for 948 that could hand an ordinary user root control. An earlier model, GPT-5.6 Sol, managed 166 with the same tooling. OpenAI says it has sent 304 critical issues to maintainers along with suggested fixes.
Mozilla classified the 81 browser bugs as high severity and has started shipping patches; OpenAI says each report came with evidence and reproducible tests. Company officials also said the model developed and validated a critical remote code execution flaw in Brotli, an open-source compression tool, and that every vulnerability was reported and patched.
On the ExploitGym benchmark, GPT-6 Cyber succeeded on 48.5% of exploitation attempts, compared with 32.1% for GPT-5.6 Cyber. OpenAI also claims the new model beats GPT-6 Astra on most cyber tasks.
Guardrails after rogue agents
Much of the pitch centers on restraint. Researcher Eric Wallace said OpenAI has sped up work on reinforcement learning and on curbing reward hacking since the summer, when its 5.6 models were tied to a string of unauthorized hacks. In August, close to 700 agents running on an OpenAI model broke out of a test sandbox and used software holes to get into Hugging Face systems.
Wallace said GPT-6 Cyber's training included a simulated exercise modeled on that episode, meant to teach the model which parts of a codebase or company it may and may not touch.
In one test, the model got an exploitation assignment next to deliberately weak services it was told to leave alone. Without deployment safety prompts, it held back 99.93% of the time, versus 94.49% for GPT-5.6 Cyber; with the prompts, it never went after the decoys.
New safeguards include monitoring that compares the model's actions with what a user is allowed to do, protections against helping attack industrial controls and other safety-critical systems, tighter identity and account requirements, and more frequent check-ins with OpenAI on high-risk requests. Industrial security firm Dragos tested Daybreak Red models in a lab and fed back findings.
The skill floor is dropping for the reconnaissance and vulnerability research needed before an attack
Rob T. Lee, Dragos chief AI officer
A separate Daybreak Red group for approved critical infrastructure defenders will allow sanctioned testing of those systems. OpenAI also says it wants to extend Daybreak to organizations of every size, independent researchers and open-source maintainers.
A dispute over user data
The launch arrives amid friction over a related program. Last week OpenAI told Daybreak Blue members that, beginning at month's end, data from their personal-account sessions would be used to improve its models, including misuse safeguards, even for users who had chosen not to share. Some users objected online, fearing their work would train models that compete with their businesses.
Fouad Matin, OpenAI's safety lead, said customers have since been told they can keep opting out of model training. He said the collected data is meant to prevent abuse of OpenAI's models, not to train on defenders' work.
