OpenAI announced GPT-6 Cyber on Nov. 5, presenting it as a cybersecurity-focused model built to help authorized researchers find and validate software vulnerabilities while remaining within the scope of an approved task. The company is not opening it to everyone: initial use is limited to a select group of approved Daybreak Red customers through the API and Codex, with expansion to other approved Daybreak Red users promised in stages. Reuters independently reported the launch and the small initial customer group the same day. [OpenAI] [Reuters]
What OpenAI announced
OpenAI’s Nov. 5 announcement describes GPT-6 Cyber as a specialized model for security work, rather than a general-purpose model receiving a new public-facing feature. Its stated purpose is to help researchers inspect software, identify weaknesses, validate whether a suspected vulnerability can be exploited, and assist in finding fixes. Those capabilities are dual-use: the same class of analysis that helps defenders test systems could be misused if aimed at systems without permission. OpenAI frames the model around authorized research and scope limits, while tying its distribution to a controlled customer program called Daybreak Red. [OpenAI]
The launch is therefore both a model announcement and an access-policy announcement. OpenAI says GPT-6 Cyber is initially available through its application programming interface and Codex to a select group of approved Daybreak Red customers. It says access will expand progressively to all Daybreak Red customers “soon”; Reuters likewise reports an expansion to all approved Daybreak Red customers over the coming weeks. That is a plan, not confirmation that every eligible organization already has access. The initial group is the only availability described as current in the announcement. [OpenAI] [Reuters]
Reuters characterized the model as designed to find and exploit security flaws, test their consequences and help fix them before attackers do. That independent account is consistent with OpenAI’s description of deeper security workflows under Daybreak Red, including exploit development and authorized testing. The key qualifier is authorization: the sources describe access for vetted security work, not a broadly available consumer tool or permission to target arbitrary systems. This article reports the launch and the claims around it without reproducing operational techniques for exploiting software. [Reuters] [OpenAI]
Benchmarks and the Critical rating
The headline benchmark is ExploitGym. OpenAI reports a 48.5% score for GPT-6 Cyber, compared with 32.1% for GPT-5.6 Cyber, a difference of 16.4 percentage points. The company describes the test as measuring whether a model can exploit a specified vulnerability in a remote target. OpenAI also characterizes the result as a 51% relative improvement. These are company-reported results on its stated evaluation; they should not be read as a broad measurement of real-world security performance or as an independently reproduced benchmark. [OpenAI]
OpenAI also assigns GPT-6 Cyber a “Critical” cybersecurity rating under its Preparedness Framework, compared with “High” for GPT-5.6 Cyber. The rating is the company’s own classification of model capability in cybersecurity, not an external certification or a government designation. The distinction matters because “Critical” can sound like a general verdict on product safety. In the announcement it refers specifically to OpenAI’s preparedness assessment of cyber capability, and is presented alongside its controls and deployment restrictions. Reuters confirms the category and the change from the prior model. [OpenAI] [Reuters]
In another evaluation, OpenAI says GPT-6 Cyber found 40.3% of vulnerabilities in an internal web-penetration test, compared with 32.0% for GPT-6 Astra. The task involved probing web applications without source-code access. The company also describes a controlled RemoteExploitBench evaluation using known vulnerabilities in test deployments of common software. It says GPT-6 Cyber completed 95% of those tasks, while GPT-6 Astra with Daybreak Blue refused all of them; OpenAI says the Cyber model’s only failure was an exploit it could not make work remotely. The comparison is specific to this evaluation design and configuration, not a claim that GPT-6 Astra is generally incapable or that real environments will yield the same results. [OpenAI]
The announcement’s benchmark figures provide useful detail about what OpenAI chose to measure: discovery, validation, and staying within task boundaries. They do not, by themselves, establish how often the model will uncover previously unknown flaws in everyday deployments, how much analyst time it will save, or how consistently its proposed fixes will be correct. OpenAI’s launch materials offer its own evaluation descriptions and point readers to a system card for further settings. Neither the benchmark score nor the framework label substitutes for independent testing across varied organizations and software. [OpenAI]
What OpenAI says it found in Linux and Firefox
OpenAI says its researchers used GPT-6 Cyber to develop 948 privilege-escalation exploits in Linux and to identify 81 previously unknown high-severity vulnerabilities in Firefox. Reuters adds detail from the company’s account: 588 of the 948 Linux vulnerabilities were previously unknown, and OpenAI disclosed 304 critical Linux findings to maintainers alongside proposed patches. The figures describe different categories in the company’s reporting and should not be collapsed into one count. “Developed exploits,” “previously unknown vulnerabilities,” and “critical findings disclosed” are not interchangeable measures. [OpenAI] [Reuters]
For Firefox, OpenAI says the 81 findings were previously unknown and high severity, and that the model helped develop candidate patches. Reuters reports that Mozilla classified the flaws as high severity and had begun releasing fixes. This gives the announcement a concrete defensive dimension: at least some reported findings moved into vendor review and remediation rather than remaining benchmark demonstrations. Still, neither source suggests that every finding has been fixed, that all patches were generated by the model, or that the model alone performed the research. OpenAI describes work by its researchers using the model, and Mozilla’s release of fixes is a separate step in the process. [OpenAI] [Reuters]
The Linux total also requires careful wording. OpenAI’s page says researchers developed 948 privilege-escalation exploits in Linux. Reuters reports that OpenAI said those corresponded to Linux kernel vulnerabilities, including 588 previously unknown cases, and separately says the company disclosed 304 critical findings to maintainers with proposed patches. These totals can overlap or use different screening and disclosure criteria; the two sources do not provide enough detail to reconcile each record. The article therefore attributes each number to its source rather than presenting them as independently audited tallies. [OpenAI] [Reuters]
Scope controls and remaining questions
OpenAI presents scope adherence as a central feature of the model, not a secondary safeguard. In its ExploitGym Honeypot evaluation, it says GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without a deployment safety prompt and in 100% with that prompt. OpenAI compares the no-prompt result with 94.49% for GPT-5.6 Cyber. The company says the evaluation places a model in an isolated environment where ancillary services are intentionally attractive targets, while the task is limited to a designated target. The reported percentages are test outcomes under those conditions, not a guarantee of perfect conduct in every deployment. [OpenAI]
The difference between 99.93% and 100% also makes clear why deployment conditions matter. OpenAI attributes the latter figure to the addition of its deployment prompt, which reinforces the allowed scope. A result in a controlled test does not establish how users will configure a system, whether real-world task boundaries will be unambiguous, or how the model will behave under every novel instruction. The announcement describes dedicated training and safety evaluations, but does not present those as eliminating misuse risk. Restricting access to approved customers is another layer of governance; it is not proof that misuse is impossible.
OpenAI says Daybreak Red is intended for more advanced security tasks, while Daybreak Blue supports verified defenders conducting vulnerability validation and fix work with fewer unnecessary refusals. It also says its general GPT-6 Astra model supports code review, vulnerability triage and patch development. These distinctions frame GPT-6 Cyber as a specialized capability inside a broader access program. The company’s approach is to offer more sensitive functionality in the more restricted tier, rather than make the same capability universally available at launch. That policy choice is relevant to the product’s reach, but customers and outside observers will still need evidence about how screening, monitoring and incident response work in practice. [OpenAI]
Who can use GPT-6 Cyber now
As of the Nov. 5 announcement, GPT-6 Cyber is available only to a select group of approved Daybreak Red customers, using the API and Codex. OpenAI’s language sets out a future expansion to all Daybreak Red customers, but does not provide a precise date in the article. Reuters reports that OpenAI expected this expansion over the coming weeks. Readers should distinguish this planned broader availability from present access: being a Daybreak customer, or applying to the program, does not necessarily mean that an organization has immediate access to GPT-6 Cyber. Approval and phased rollout remain part of the stated access model. [OpenAI] [Reuters]
OpenAI separately says GPT-5.6 Cyber is becoming available to all Daybreak Red customers. That broader availability applies to the earlier model, not GPT-6 Cyber. The company also describes expansion of Daybreak Blue alpha access for GPT-6 Astra to more organizations and individual security researchers. These are three different access statements involving distinct model and program combinations. Treating them as one general opening of GPT-6 Cyber would overstate what is available. Reuters similarly distinguishes the current select-customer launch from the planned GPT-6 Cyber expansion and the immediate availability of GPT-5.6 Cyber to Daybreak Red users. [OpenAI] [Reuters]
For organizations that do not qualify for Daybreak Red, OpenAI points to general models and defensive tools for tasks such as code review and patch development. Its announcement invites new users and organizations to apply for Daybreak, but it does not spell out eligibility standards, review timelines, customer pricing, or the number of organizations in the first cohort. Reuters calls the initial group small and vetted. Without additional public details, the launch signals a gated route to access rather than a predictable self-service signup process. [OpenAI] [Reuters]
Why the launch matters
GPT-6 Cyber arrives amid a broader debate over whether increasingly capable AI systems can help defenders find weaknesses faster than attackers can exploit them. OpenAI argues that security teams have a limited window to use frontier models to secure software, and says the new system has already contributed to work on Linux, Firefox and its own infrastructure. Those are the company’s rationale and reported examples, not a neutral forecast that defenders will necessarily gain the advantage. The same capabilities that make vulnerability discovery useful to defenders also explain why OpenAI has tied access to a vetted program. [OpenAI]
The policy challenge is not resolved by describing a model as “aligned” or “critical.” A controlled product can widen access for legitimate security teams while increasing the importance of careful authorization, monitoring and coordinated disclosure. OpenAI’s stated plan to expand to all approved Daybreak Red customers means that the gate may broaden, but the sources do not indicate unrestricted public access. Its reported scope-adherence score is encouraging as a test result, but practical safety will depend on product controls and customer practices as well as model behavior. Those operational details will be important to assess as the staged rollout proceeds.
This launch is distinct from OpenAI’s GPT-6.1 Terra announcement. Terra was presented as a general professional model with visual reasoning and computer-use features; GPT-6 Cyber is a cybersecurity-specific model, classified by OpenAI as Critical under its cyber Preparedness Framework and restricted at launch to selected Daybreak Red customers. The difference is not simply a new benchmark or a renamed capability. It is a separate product announcement with a separate safety profile, intended use and distribution plan. [OpenAI]
The next evidence to watch is practical rather than rhetorical: whether OpenAI expands access on the timetable it outlined, how maintainers process disclosed findings, and whether external evaluations reproduce the company’s performance and scope-control claims. For now, the defensible summary is narrower. OpenAI launched a cyber-focused model on Nov. 5; Reuters corroborated the date, rating and restricted initial availability; and the headline performance and vulnerability counts remain claims attributed to OpenAI. [OpenAI] [Reuters]
FAQs
What is GPT-6 Cyber?
OpenAI describes GPT-6 Cyber as a cybersecurity-focused model for authorized vulnerability research, validation and related defensive work. The company says it is designed to operate within a researcher-defined scope. [OpenAI]
What did it score on ExploitGym?
OpenAI reports 48.5% for GPT-6 Cyber and 32.1% for GPT-5.6 Cyber. These are company-reported benchmark scores, not independently verified results in the cited Reuters report. [OpenAI]
Why is its cybersecurity rating called Critical?
“Critical” is OpenAI’s Preparedness Framework category for the model’s cybersecurity capability. OpenAI says GPT-5.6 Cyber was rated High; Reuters independently reports the new rating and comparison. It is not described as an external certification. [OpenAI] [Reuters]
Is GPT-6 Cyber generally available?
No. OpenAI says initial access is for a select group of approved Daybreak Red customers through API and Codex. Broader access to approved Daybreak Red customers is planned, but the Nov. 5 announcement does not say that rollout is complete. [OpenAI] [Reuters]
What Linux and Firefox findings did OpenAI report?
OpenAI says its researchers developed 948 privilege-escalation exploits in Linux and identified 81 previously unknown high-severity Firefox vulnerabilities. Reuters attributes additional counts to OpenAI: 588 of the Linux vulnerabilities were previously unknown, and 304 critical findings were disclosed to maintainers with proposed patches. The categories and totals should not be treated as interchangeable. [OpenAI] [Reuters]