OpenAI announced GPT-6 Cyber on 5 November 2026, a specialized model built for one job: finding software vulnerabilities, proving they can be exploited, and helping security teams fix them. It is a variant of the flagship GPT-6 Astra, given extra training for security research — and OpenAI's own numbers say the jump from its predecessor is large: 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, which the company frames as a 51% relative improvement. The model already has real-world results behind it. It powered the Linux kernel research OpenAI announced the day before, and it found 81 previously unknown high-severity vulnerabilities in Firefox — software that had already been tested extensively by frontier models.
There is a catch that matters more than the benchmarks: almost nobody reading this can use it. GPT-6 Cyber is available only through OpenAI's API and Codex to a select group of approved Daybreak Red customers, with access expanding to all Daybreak Red customers "soon". It is a launch, but a gated one.
What the model actually does
GPT-6 Cyber is not a general chat model. It is trained to understand code, find vulnerabilities, write working exploits, and help draft patches — while staying inside the scope a researcher is authorized to test. Three evaluations in OpenAI's announcement define it:
- ExploitGym — testing whether a model can exploit a specified vulnerability in a remote target: GPT-6 Cyber scores 48.5%, about 16 percentage points above GPT-5.6 Cyber's 32.1%.
- Web penetration testing — probing web applications without access to their source code: 40.3% recall of vulnerabilities, against 32.0% for GPT-6 Astra.
- RemoteExploitBench — a new benchmark built on known vulnerabilities in applications such as GitLab, PostgreSQL, and Jenkins in a controlled environment: GPT-6 Astra running under Daybreak Blue refused 100% of the tasks, while GPT-6 Cyber completed 95%; its only failure was an exploit it could not get working remotely.
The third number is the point. Daybreak is OpenAI's access program for cybersecurity work: Blue is for verified defenders doing vulnerability validation with fewer refusals, Red is for deeper exploit development and authorized security testing. The gap between Astra refusing everything and Cyber completing nearly everything is the difference between a general model with safety brakes and a specialist built to do the work.
OpenAI also made alignment — staying within authorized scope — a central part of the release. In its ExploitGym Honeypot evaluation, where deliberately easy-to-exploit supporting services tempt the model outside its task, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without its deployment safety prompt, and 100% with it, up from 94.49% for GPT-5.6 Cyber. That is a company-reported evaluation, not an independent result, but the direction is the story OpenAI wants told: a more capable offensive model that is also better at following the rules of engagement.
The real-world results
Benchmarks are one thing; the model has already been put to work on codebases that matter:
- Linux kernel: 948 working exploits. OpenAI's researchers used GPT-6 Cyber to reproduce 1,724 vulnerabilities and develop working local privilege-escalation exploits for 948 of them — more than five times the 166 exploited by GPT-5.6 Sol using the same research harness. Following the Linux community's guidance, 304 critical findings were disclosed to maintainers with candidate patches.
- Firefox: 81 novel high-severity flaws. During a vulnerability research campaign, the model produced evidence and reproducible tests for 81 vulnerabilities Mozilla classified as high severity, and showed that a further 30 known issues were more severe than previously reported. Mozilla began landing fixes shortly after receiving the reports; its Firefox 159 security advisory credits OpenAI researchers for some of the fixes.
- Brotli: a critical RCE. OpenAI's security researchers used GPT-6 Cyber alongside other models to develop and validate exploits for a critical vulnerability in Brotli, a compression library used by Node.js and other software, showing how the flaw could enable remote code execution in downstream applications.
Security teams at Cisco, Calif, and CodeRabbit are already using the model for vulnerability research and code review, and OpenAI says it has used it to secure its own infrastructure — including red-teaming campaigns in which agents chained vulnerabilities to escape microVMs and execute code on underlying hosts.
The "Critical" rating — and the safeguards around it
Under OpenAI's Preparedness Framework, GPT-6 Cyber is classified as Critical for cybersecurity capabilities, up from High for GPT-5.6 Cyber. It is the company's highest capability tier for this risk area, and the rating is the reason access is gated rather than open.
The safeguards announced alongside the model are correspondingly heavy: cyber-specific monitoring for whether actions match the user's authorization, targeted monitoring around emerging threats (for example, during the coordinated disclosure of critical PHP vulnerabilities), Private Safety Processing for reviewing related interactions, and IP allowlisting by default for Daybreak Red API access. For critical infrastructure, OpenAI is adding a new Daybreak Red for Critical Infrastructure group for approved defenders, backed by a $1 billion Daybreak for Frontline Defenders grant program for resource-constrained teams, and it partnered with Dragos to evaluate the model in a controlled lab environment for operational technology scenarios.
None of this is independently verified yet — the numbers come from OpenAI and the evaluators it worked with, and the system card is the place to check the full evaluation settings. But the structure is notable: OpenAI is treating a more capable security model as something that needs monitoring infrastructure, not just a release note.
Access and price
GPT-6 Cyber is initially available through the API and Codex to selected approved Daybreak Red customers. OpenAI says it will progressively expand access to all Daybreak Red customers, and is also making GPT-5.6 Cyber available across that tier. Daybreak Blue alpha access for GPT-6 Astra is expanding to more organizations and individual security researchers.
Pricing, as listed on OpenAI's model pages and reported by VentureBeat: $20 per million input tokens and $100 per million output tokens, with cached input at $2 per million. That is about 60% more for input and a third more for output than GPT-5.6 Cyber ($12.50 / $75), and well above competing cyber models from Anthropic and Google at published rates — a price that matches the positioning: a specialist tool for approved security teams, not a general-purpose model.
Why it matters even if you can't use it
For most readers, GPT-6 Cyber changes nothing directly — you cannot call it without going through OpenAI's vetting. It matters for a different reason: the software you use every day is getting patched because of it. Eighty-one Firefox flaws and hundreds of Linux kernel exploits, found by a model and disclosed with candidate patches, are a concrete security dividend that lands in browsers and servers whether or not you ever touch the model.
The other reason it matters is the tension OpenAI states openly. In its announcement, the company says there is "a limited window to use frontier models to secure software" — the same capabilities that find flaws for defenders can be misused by attackers, which is why the model ships behind an approval wall, IP allowlisting, and monitoring. The launch is therefore two stories at once: AI is now good enough to find vulnerabilities at a scale humans cannot match, and the labs know that capability is dangerous enough to keep behind a gate. How well that gate holds is the question the next year will answer.
Sources and classification
Classification: Official Announcement (OpenAI, 5 November 2026). Launch facts, benchmark figures, and the real-world vulnerability counts are from OpenAI's announcement and system card; all capability numbers are OpenAI's own, with no independent verification published yet. Pricing is from OpenAI's model pages as reported by VentureBeat (checked 5 November 2026). Checked 6 November 2026.
- OpenAI, "GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents" (official announcement, 5 November 2026): https://openai.com/index/introducing-gpt-6-cyber/
- OpenAI, GPT-6 Cyber system card: https://deploymentsafety.openai.com/gpt-6-cyber
- Mozilla, Firefox 159 security advisory (fixes credited to OpenAI researchers, 27 October 2026): https://www.mozilla.org/en-US/security/advisories/mfsa2026-118/
- VentureBeat, "OpenAI launches GPT-6 Cyber with stronger exploit research and higher API price than Anthropic and Google" (5 November 2026) — pricing comparison and access details: https://venturebeat.com/technology/openai-launches-gpt-6-cyber-with-stronger-exploit-research-and-higher-api-price-than-anthropic-and-google
- Neowin, "OpenAI announces GPT-6 Cyber, its most capable cybersecurity model yet" (5 November 2026) — launch context and the Linux kernel research: https://www.neowin.net/news/openai-announces-gpt-6-cyber-its-most-capable-cybersecurity-model-yet/
- Unite.AI, "OpenAI Launches GPT-6 Cyber With Restricted Access and Scope Controls" (5 November 2026) — access tiers and Mozilla advisory corroboration: https://www.unite.ai/openai-launches-gpt-6-cyber-with-restricted-access-and-scope-controls/