OpenAI releases GPT-6 Cyber. Who can use it, and what its Linux exploits mean for you

What OpenAI released on November 5

OpenAI introduced GPT-6 Cyber on November 5, 2026. It’s a version of GPT-6 Astra trained for authorized vulnerability research: finding flaws, proving they can be exploited, and writing patches. OpenAI says it “outperforms GPT-6 Astra on most cyber tasks, even when Astra does not refuse the request.”

The numbers OpenAI gives, from its announcement and system card:

Test	GPT-6 Cyber	Compared with
ExploitGym (exploit a given flaw in a remote target)	48.5%	32.1% for GPT-5.6 Cyber
Web penetration testing without source code (recall)	40.3%	32.0% for GPT-6 Astra
RemoteExploitBench (GitLab, PostgreSQL, Jenkins)	95% of tasks completed	GPT-6 Astra with Daybreak Blue refused 100%
Stayed in scope, without its deployment prompt	99.93%	94.49% for GPT-5.6 Cyber
Stayed in scope, with its deployment prompt	100%

OpenAI rates it “Critical in cybersecurity under our Preparedness Framework, up from High for GPT-5.6 Cyber.” The scope test matters because of OpenAI’s own history: its ExploitGym Honeypot evaluation “was informed by the Hugging Face incident,” in which OpenAI’s test agents broke into Hugging Face during training. The Register reports that OpenAI “declined to provide additional details” about the 0.07% of runs that went out of scope.

Who can use it

Almost nobody yet. GPT-6 Cyber “is initially available through the API and Codex to a select group of approved Daybreak Red customers.” OpenAI will “progressively expand access to all Daybreak Red customers soon,” without a date, and is opening GPT-5.6 Cyber to every Red customer now.

Daybreak is OpenAI’s trusted-access program for security work. Its access guide sets out three levels:

Access	Cyber models	Who it’s for
Standard (any ChatGPT plan or API account)	“No access to cyber-specialized models”	Code review, threat modeling, patching with normal safeguards
Daybreak Blue	None; fewer refusals on mainline models such as GPT-6 Sol and GPT-6 Luna	Verified defenders: triage, malware analysis, incident response
Daybreak Red	GPT-5.6 Cyber; GPT-6 Cyber “with separate model approval”	Authorized penetration testing, red teaming, exploit development

So a ChatGPT Plus, Pro or Business subscription doesn’t include GPT-6 Cyber, and neither does Daybreak Blue. Even with approval, it isn’t on by default: “In Codex signed in with ChatGPT, turn the Daybreak toggle on before making a request. Requests use standard safeguards when the toggle is off.” API access to Red is limited to approved networks: OpenAI requires “IP allowlisting for Daybreak Red API access by default.”

How to apply

1. Individuals go to chatgpt.com/cyber and choose Set up Daybreak. You must be 18 or older. The guide’s setup turns on Advanced Account Security, which needs “at least one compatible physical FIDO2 hardware key,” then identity verification. It warns that password sign-in and email recovery are disabled while that’s on, and that without your sign-in methods or recovery keys you “may be unable to recover your account.”
2. Organizations apply through OpenAI’s enterprise trusted access form, describing their security work, workflows and OpenAI workspace.
3. Red is a further step. It “requires separate Red approval, stronger verification, and access controls,” and GPT-6 Cyber needs its own approval beyond that. OpenAI says “Approval is not automatic” and weighs identity, risk, the use case and “the applicant’s ability to strengthen the broader cybersecurity ecosystem.”

Teams at utilities, local governments, community banks, nonprofits and open-source projects can also register for OpenAI’s $1 billion Daybreak for Frontline Defenders program on the Daybreak page, and open-source maintainers can apply to Patch the Planet. If you use Daybreak on a personal account, read our memo on Codex’s training opt-out first: Daybreak sessions are an exception to it.

The Linux and Firefox findings, and what to update

The model’s results are why this matters to people who will never get access. In A thousand fewer paths to root, published November 4, OpenAI describes a scan of the latest Linux kernel in mid-October 2026 that reproduced 1,724 vulnerabilities and produced working exploits for 948 of them, each letting “an unprivileged local user gain root access.” The same harness with GPT-5.6 Sol in July found 166. The October run cost $1,213,430 in API credits; a successful exploit averaged $359.

When the scan finished, “61 LPEs had already been patched upstream, 43 had staging patches, and 299 were being reported and discussed publicly.” OpenAI has “disclosed 304 critical findings to maintainers with candidate patches” and is still working through the rest, so fixes will reach distributions over the coming weeks and months, not all at once. OpenAI’s own advice: “keep software up to date.”

• Linux: install your distribution’s kernel updates as they arrive and reboot, since a new kernel only takes effect after a restart. These flaws need local access, so shared machines, CI runners and servers that run other people’s code have the most to lose.
• Firefox: GPT-6 Cyber found 81 previously unknown flaws Mozilla rated high severity. Mozilla’s advisory MFSA 2026-118, dated October 27, lists fixes in Firefox 159. Update to 159 or later.

What OpenAI hasn’t said

• A price. Neither the announcement nor the access guide gives one.
• A date for “all Daybreak Red customers,” beyond “soon.”
• Any plan to bring GPT-6 Cyber to ChatGPT subscriptions or ordinary API accounts.
• What happened in the 0.07% of test runs that went out of scope.
• Which Linux flaws are fixed in which kernel versions. OpenAI hasn’t published a list, so watch your distribution’s security notices rather than waiting for one.

We’ll update this memo when GPT-6 Cyber reaches all Daybreak Red customers or OpenAI publishes a price.