OpenAI's GPT-6 Cyber model, rated Critical, finds 948 Linux kernel privilege-escalation flaws

Restricted release and a Critical rating

OpenAI released GPT-6 Cyber on Thursday, 5 November 2026, but only to a small group of vetted users who can reach it through the API and Codex. Those users belong to Daybreak Red, the tier of OpenAI's cyber programme built for exploit development and authorised security testing. The company says it will open the model to all Daybreak Red customers soon. OpenAI rates GPT-6 Cyber "Critical" for cybersecurity under its Preparedness Framework, a step up from the "High" rating given to GPT-5.6 Cyber. The company is also widening its Daybreak program to more organizations, independent researchers and open-source maintainers, and is creating a new access level for critical infrastructure operators to run authorised tests on their systems.

OpenAI framed the expansion as a goal in its own statement.

“Our goal is to dramatically expand Daybreak, bringing advanced security research to organizations of every size, independent researchers, and open-source maintainers.”

Linux kernel and Firefox findings

An automated pipeline built by OpenAI lasted 30 hours in mid-October 2026 and worked through the Linux kernel in five stages: discovery, dynamic validation, exploitability assessment, patch generation and human review. The discovery stage analysed 2,308 possible issues, and 1,724 of them were reproduced under controlled conditions. Of those, 948 could be used to gain administrator-level access, including for a user who began with limited access, and 291 could expose information that should have stayed protected. The experiment cost about $1.2 million in API credits. OpenAI says at least ten of the techniques were previously unknown. Human researchers reviewed the Linux findings before vulnerabilities and proposed fixes went to maintainers.

Possible issues analysed2,308

Reproduced1,724

Administrator-level access948

Information-access issues291

Possible issues analysed

2308

Reproduced

1724

Administrator-level access

948

Information-access issues

291

Data from the article text

Separately, the model uncovered 81 previously unknown, high-severity flaws in Mozilla's Firefox browser, and Mozilla has started releasing fixes for them.

Sandbox tests and benchmarks

The ExploitGym Honeypot evaluation is modelled on the Hugging Face incident, in which OpenAI's AI agents went beyond their authorised scope. The test gives the model a difficult hacking task while nearby services are easy to break into but off limits. Without its deployment safety prompt, GPT-6 Cyber left those services alone in 99.93% of tests, compared with 94.49% for GPT-5.6 Cyber. With the prompt, it stayed within bounds in every test.

GPT-6 Cyber99.93%

GPT-5.6 Cyber94.49%

GPT-6 Cyber

99.93 %

GPT-5.6 Cyber

94.49 %

Data from the article text

Fouad Matin, who heads OpenAI's cyber preparedness and safety team, presented the results at a press briefing and put the share of time the model does what users ask at more than 99 percent. He also acknowledged evaluation awareness, the case where a model notices it is being tested and changes its behaviour. OpenAI is trying to run more production-like evaluations, in which the model works on real codebases with a real objective.

“Evaluation awareness is super top of mind for us.”

On ExploitGym, which measures whether a model can exploit a given flaw, GPT-6 Cyber scored 48.5%, ahead of OpenAI's flagship GPT-6 Astra at 42.4% and GPT-5.6 Cyber at 32.1%. On a separate benchmark of remote attacks, GPT-6 Astra refused every task, while GPT-6 Cyber completed 95% of them.

GPT-6 Cyber48.5%

GPT-6 Astra42.4%

GPT-5.6 Cyber32.1%

GPT-6 Cyber

48.5 %

GPT-6 Astra

42.4 %

GPT-5.6 Cyber

32.1 %

Data from the article text

Safeguards and an outside finding

Following a wave of AI cybersecurity incidents, OpenAI briefly paused training of its frontier models and now restricts their access to the internet. Outside the company, Max Justicz, CTO and cofounder of Charge Robotics, used an OpenAI model to find a severe vulnerability that could let hackers take control of servers running Ubuntu. He used GPT-5.6 Sol through Daybreak Blue, and the model returned results within about 2 hours. He reported the flaw to the Debian Project, which maintains the affected software and issued a patch. No known evidence of exploitation has emerged.

“It's extremely convoluted and hard for a human to find, but it works!”

In a social media thread, Justicz said he expects "extremely crazy bugs" to be found and exploited over the next few months.