OpenAI Launches GPT-6 Cyber, Claims 'Critical' Rating and 948 Linux Kernel Exploits

OpenAI has released GPT-6 Cyber, a specialized cybersecurity model it classifies as "Critical" under its own Preparedness Framework. The company says the model developed working exploits for 948 Linux kernel local privilege escalation vulnerabilities and found 81 previously unknown high-severity flaws in Firefox. Access is restricted to approved "Daybreak Red" customers.

Linux kernel campaign: 948 exploits at $1.2 million

In a mid-October run against the latest Linux kernel master branch, OpenAI's vulnerability research pipeline—using what it calls an "upcoming cyber model"—dynamically reproduced 1,724 vulnerabilities and generated 948 working local privilege escalation exploits. The company reports an API cost of $1,213,430, or roughly $1,279 per exploit [1].

The pipeline operates in five stages: scout agents identify audit targets, analyst agents scan for candidate findings, triage agents validate findings in isolated VMs with Kernel Address Sanitizer, exploit-development agents test exploitability against standard mitigations (SMEP, SMAP, KASLR, HARDENED_USERCOPY), and human researchers review results before reporting to maintainers [1].

OpenAI says 588 of the 948 Linux exploits were previously unknown [2][5]. The company applied historical Pwn2Own ($20,000) and TyphoonPWN ($70,000) bounty values to estimate a "hypothetical valuation" of $18.9M–$66.3M, or 16x–55x the pipeline cost [1]. On Google's hardened kernelCTF environment, the same pipeline found 21 LPE exploits at a cost of $32,936, compared to 26 confirmed 0-day LPEs accepted by kernelCTF in all of 2026 [1].

During the 10-day scan, OpenAI observed Linux maintainers patching 61 LPEs upstream, staging 43 more, and publicly discussing 299 others [1]. The company says it has disclosed 304 critical findings with candidate patches and is working with "Patch the Planet" partners on the remainder [3].

Firefox findings and novel techniques

OpenAI reports 81 novel high-severity Firefox vulnerabilities, with Mozilla releasing fixes for many [3]. In a sample of 166 exploits, researchers identified 10 exploitation techniques they say were new to external kernel experts: two persistent elastic objects for kernel cache spraying, two page grooming techniques, five data-only techniques for privilege escalation or arbitrary write, and one arbitrary type confusion technique [1]. Technical details were withheld "for safety reasons" [1].

Benchmarks and alignment claims

On OpenAI's ExploitGym benchmark, GPT-6 Cyber scores 48.5% versus 32.1% for GPT-5.6 Cyber [3]. On a new "RemoteExploitBench" testing known vulnerabilities in GitLab, PostgreSQL, and Jenkins, GPT-6 Cyber completed 95% of tasks while GPT-6 Astra with Daybreak Blue safeguards refused 100% [3].

For scope adherence, OpenAI's ExploitGym Honeypot evaluation—designed after the Hugging Face incident—shows GPT-6 Cyber avoided out-of-scope services in 99.93% of cases without deployment prompts, and 100% with them, up from 94.49% for GPT-5.6 Cyber [3]. The company cites chain-of-thought examples where the model reasons "Can't assume extra API or attack controller not target" and "don't explore out of scope credentials" [3].

Restricted access and the Daybreak program

GPT-6 Cyber is initially available only to a select group of approved Daybreak Red customers via API and Codex, with expansion to all Daybreak Red customers planned over "coming weeks" [3]. Daybreak Blue, for defensive tasks, is also expanding alpha access for GPT-6 Astra [3]. OpenAI says "thousands of defenders across 2,000 approved organizations and workspaces already use Daybreak" [6].

Chris Hughes of Noma Security notes that even 0.07% out-of-scope actions "at the scale these models operate... means actions against systems nobody authorized" [2]. Max Justicz, CTO of Charge Robotics, told TechCrunch he found a severe Ubuntu vulnerability using GPT-5.6 Sol through Daybreak Blue, but says the tier was recently "nerfed" to be less useful for vulnerability finding [4]. Justicz spent about $5,000 on subscriptions and tokens to find several exploits [4].

Remediation gap and resource questions

Hughes points out that OpenAI's Patch the Planet dashboard shows "a less than 25% remediation rate thus far," and warns that improved finding capability "is likely to exacerbate that further, meaning more things found but unfixed and likely to be found by malicious actors" [2]. OpenAI provided no measurements showing how much its models accelerate enterprise patch deployment or reduce exposure [2].

TechCrunch notes the Linux bug hunt cost $1.2 million—significantly less than the approximately $6 million OpenAI invested in a fluid dynamics problem—and questions whether the company is devoting enough frontier compute to the "narrow window" it describes for defenders [4]. OpenAI researcher Eric Wallace said the company is "allocating a significant amount of our frontier compute towards discovering and patching vulnerabilities in open source codebases, fully pro bono" [4].

Critical infrastructure push

Alongside GPT-6 Cyber, OpenAI announced "Daybreak for Frontline Defenders," a $1 billion commitment to subsidize access for resource-constrained U.S. organizations—water utilities, electric grid operators, state and local governments, community banks, nonprofits, and open-source maintainers—targeting consumption over six months [6]. A pilot with the Multi-State Information Sharing and Analysis Center (MS-ISAC) will pair Daybreak access with guided training for public sector and water system defenders [6]. Over 35 partner products in the "Daybreak Defense Network" aim to integrate the models into existing defender workflows [6].

What remains unverified

The Linux and Firefox results come from OpenAI's own research pipeline and disclosures; independent reproduction has not been reported. The "Critical" classification is under OpenAI's internal Preparedness Framework. Benchmark scores (ExploitGym, RemoteExploitBench, ExploitGym Honeypot) are from OpenAI's evaluations. The 10 novel exploitation techniques were withheld from publication. The $1 billion Daybreak commitment is a pledge with implementation underway but not yet independently assessed.