OpenAI launches GPT-6 Cyber: 948 Linux exploits for $1.2M compute

OpenAI introduced GPT-6 Cyber on 5 November 2026. OpenAI describes it as a cybersecurity model for finding exploitable vulnerabilities and developing fixes, trained to stay within authorised scope. It is available first through the API and Codex to a select group of customers of Daybreak Red, OpenAI's programme for offensive security researchers, according to The Decoder's report of the announcement.

What OpenAI reports

All numbers here come from OpenAI, as relayed by The Decoder. On ExploitGym, which tests whether a model can exploit a given vulnerability in a remote system, GPT-6 Cyber scores 48.5%, against 32.1% for GPT-5.6 Cyber. On a new test for finding flaws in web apps without source code, OpenAI reports 40.3%, against 32.0% for the standard GPT-6 Astra.

OpenAI rates the model "Critical" in the cybersecurity category of its own Preparedness Framework, the same rating it gave GPT-6 Astra. The Decoder notes that no independent testing has been published.

Real targets and real cost

OpenAI says a targeted campaign found 81 previously unknown severe vulnerabilities in Firefox, even though the browser had been tested with earlier AI models, and showed 30 known issues were more severe than thought. The Decoder reports that Mozilla accepted the reports and has shipped fixes.

For the Linux kernel, OpenAI says scout and analyst agents scanned about 5,034 security-relevant locations, filtered them to 2,308 candidates, confirmed 1,724 vulnerabilities in isolated virtual machines and built working exploits for 948. The run cost $1,213,430 in compute, about $1,280 per exploit. OpenAI says 304 critical findings went to Linux maintainers with suggested fixes. TechCrunch puts the cost at $1.2 million as well.

Guardrails and access

OpenAI says it trained the model against a honeypot inspired by the Hugging Face incident. Without special safety instructions it stayed within bounds in 99.93% of cases, up from 94.49% for its predecessor; with instructions, 100%. Listed countermeasures include checks that actions match user permissions, monitoring for newly disclosed vulnerabilities, IP-based restrictions for Daybreak Red, and a recommendation to run the model in isolated environments with human oversight.

GPT-5.6 Cyber opens to all Daybreak Red users at the same time, and Daybreak Blue alpha access for GPT-6 Astra widens. OpenAI also points to a $1 billion funding programme for defenders and a subgroup for critical infrastructure testing.

The catch

TechCrunch reports that Max Justicz, a startup CTO, used an earlier OpenAI model, GPT 5.6 Sol, through Daybreak Blue to find a severe Ubuntu-related vulnerability in about two hours, spending roughly $5,000 on subscriptions and tokens across several exploits. He also says Daybreak Blue appears to have been restricted since. The Decoder adds that Epoch AI researchers showed an open-weight model, GLM, could exploit known but unpatched Firefox flaws, which is the argument against limiting access alone.

What to do: if you run patching or vulnerability management, assume discovery will outpace fixes and check how fast your team can triage a surge of reports. If you want access, the route is a Daybreak application; OpenAI has not published a price in the material I read.