OpenAI Launches GPT-6 Cyber: Critical-Rated Vulnerability Hunting, Open Only to Vetted Security Teams

How it unfolded

The model debuted on Nov 5 on OpenAI's official blog as a cybersecurity-specialized version of GPT-6 Astra. It is trained to find exploitable vulnerabilities, validate their impact and help fix them, while staying within the scope researchers set. It is available through the API and Codex, first only to a select group of approved Daybreak Red customers, with OpenAI saying access will expand to all customers in that tier over the coming weeks; GPT-5.6 Cyber opened to all Daybreak Red customers the same day, and alpha access to GPT-6 Astra in the Daybreak Blue tier is widening in parallel.

The capability figures come from OpenAI's system card. ExploitGym asks a model to turn a specified vulnerability into a working exploit against a remote target; GPT-6 Cyber scores 48.5%, above 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra, a 51% relative gain over the previous generation. In a new web penetration-testing evaluation that gives the model no application source code, GPT-6 Cyber found 40.3% of vulnerabilities, against 32.0% for Astra. On the new RemoteExploitBench, which covers known vulnerabilities in applications such as GitLab, PostgreSQL and Jenkins, Astra under Daybreak Blue refused every task while GPT-6 Cyber completed 95%, its only failure an exploit it could not reproduce remotely. Under OpenAI's own Preparedness Framework, the model is rated Critical for cybersecurity, the highest tier, up from High for GPT-5.6 Cyber and matching Astra. All figures above are company-reported; the system card carries the full results and evaluation settings.

The field record came before the model itself. OpenAI says its researchers used GPT-6 Cyber to find 81 previously unknown high-severity Firefox vulnerabilities and to show that 30 known issues were worse than reported, supplying reproducible tests and candidate patches, and that Mozilla has begun landing fixes; Mozilla's Firefox 159 security advisory of Oct 27 credits 41 high-impact fixes to OpenAI researcher Amy Burnett and OpenAI Codex Security. The same model powered the Linux kernel research published Nov 4, 'A thousand fewer paths to root': a five-stage pipeline produced 2,308 candidate findings in 12 hours, reproduced 1,724 vulnerabilities in 30 hours, and built 1,169 working local privilege-escalation exploits, corresponding after variant analysis to 948 distinct vulnerabilities, with candidate patches written for all 948; the same pipeline produced 166 with GPT-5.6 Sol in July and 24 with GPT-5.5 Cyber in June. Following the Linux community's guidance, 304 critical findings have been disclosed to maintainers with candidate patches; 10 days after the scan finished, 61 had been fixed upstream, 43 had staging patches and 299 were being reported and discussed publicly. The run cost $1,213,430 in API credits, about $1,280 per distinct vulnerability across the 948.

Alignment is the release's other thread. The ExploitGym Honeypot evaluation grew out of the Hugging Face incident: the model gets a difficult exploitation task in an isolated environment, with deliberately vulnerable supporting services beside it, and attacking those services would fall outside the task's scope. Without deployment safety prompts, GPT-6 Cyber went outside its authorized scope in 0.07% of cases, against 5.51% for the previous generation; with the prompts, it stayed within scope in every tested case. According to reports, OpenAI has not described what happened in the 0.07% of out-of-scope cases. The model also took part in two validations: researchers used it with other OpenAI models to build and validate remote-code-execution exploits for a critical flaw in the Brotli compression library, and in red-team exercises agents using GPT-6 Cyber chained vulnerabilities to escape microVMs and execute code on the underlying hosts, with the findings used to patch software, tighten network restrictions and report issues to maintainers and vendors.

What was said

OpenAI opened its blog post by framing the release as progress on both capability and alignment, saying the model is already helping secure real software, uncovering and validating thousands of exploits in security-critical codebases such as Linux and Firefox.

We're introducing GPT‑6 Cyber, a major leap in security research: more capable at finding and fixing vulnerabilities, and far more reliable staying within the boundaries researchers set.

— OpenAI official blog

The post also says there is a limited window to use frontier models to secure software, and that OpenAI's goal is to dramatically expand Daybreak, bringing advanced security research to organizations of every size, independent researchers and open-source maintainers.

It is clear that we have a limited window to use frontier models to secure software.

— OpenAI official blog

Among early users, CodeRabbit CEO Harjot Gill said in the announcement that GPT‑6 Cyber is the strongest model the company has tested for security review; VentureBeat reports CodeRabbit found 66% more medium- and high-severity issues on its internal benchmark than its production model found. OpenAI's blog also names security teams at Cisco and Calif as users of the model.

GPT‑6 Cyber is the strongest model we've tested for security review.

— Harjot Gill, CEO of CodeRabbit

On Mozilla's side, OpenAI's blog says Mozilla was supportive of the work and began landing fixes shortly after receiving the reports; Mozilla's Oct 29 blog discusses the 'patch gap' in front of automated attackers, the interval between a fix landing in code and the update reaching most users. OpenAI also says it partnered with Dragos to evaluate the model in a controlled lab for defending operational technology systems, refining safeguards from the feedback. One point remains unanswered: the details of the 0.07% out-of-scope cases — according to reports, OpenAI has not described what happened in them or whether any involved real security incidents.

Key facts

1. Publisher and date: OpenAI, Nov 5, 2026; a cybersecurity-specialized version of GPT-6 Astra
2. ExploitGym: 48.5%, above 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra; a 51% relative gain over the previous generation (company-reported)
3. Web penetration test: Probing web applications without source code, vulnerability recall of 40.3%, against 32.0% for Astra
4. RemoteExploitBench: Astra under Daybreak Blue refused every task, while GPT-6 Cyber completed 95%
5. Rating: Critical for cybersecurity under the Preparedness Framework, the highest tier, up from High for GPT-5.6 Cyber and matching Astra
6. Out-of-scope rate: 0.07% in the ExploitGym Honeypot evaluation, against 5.51% for the previous generation; 0% with deployment safety prompts
7. Firefox field record: 81 previously unknown high-severity vulnerabilities, and 30 known issues worse than reported; Mozilla's Firefox 159 advisory credits 41 high-impact fixes to OpenAI researchers
8. Linux kernel field record: 2,308 candidate findings, 1,724 reproduced, 1,169 working local privilege-escalation exploits corresponding to 948 distinct vulnerabilities; 304 critical findings disclosed to maintainers with candidate patches
9. Cost: The Linux run cost $1,213,430 in API credits, about $1,280 per distinct vulnerability across the 948; the July run with GPT-5.6 Sol produced 166 for $513,690
10. Access: First through the API and Codex to a select group of approved Daybreak Red customers, expanding to all customers in that tier over the coming weeks; applicants must pass identity and eligibility checks, and organizational customers must meet conditions including single sign-on, multi-factor authentication, managed devices and API IP allowlisting
11. Pricing: $20 per million input tokens, $100 per million output tokens and $2 cached input, above GPT-5.6 Cyber at $12.50/$75/$1.25; VentureBeat reports it above Anthropic's and Google's rival cybersecurity model pricing
12. Support: A $1 billion Daybreak for Frontline Defenders grant program subsidizes access, and a new Daybreak Red for Critical Infrastructure group is being formed; VentureBeat reports Daybreak had enrolled more than 2,000 entities as of last month