OpenAI on 5 November introduced GPT-6 Cyber, a security model it says scores 48.5% on ExploitGym, a test of whether a model can exploit a specified vulnerability in a remote target, up from 32.1% for GPT-5.6 Cyber. OpenAI says it classifies the model as “Critical in cybersecurity under our Preparedness Framework, up from High for GPT-5.6 Cyber.”
The model is “initially available through the API and Codex to a select group of approved Daybreak Red customers,” OpenAI says, with access to all Daybreak Red customers to follow “soon”. OpenAI’s post does not give a price.
Business Pill · The Model and the Harness
A short Business Engineer explainer on the difference between an AI model and the harness around it. It teaches the general idea only and says nothing about the people in this piece.
The key insight:OpenAI says GPT-6 Cyber scores 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber, and rates it Critical for cybersecurity under its Preparedness Framework, up from High. OpenAI also says the same model was behind the 948 Linux exploits it published earlier: by our arithmetic, 5.7 times the 166 that GPT-5.6 Sol managed in the same research harness.
ExploitGym: 48.5% for GPT-6 Cyber against 32.1% for GPT-5.6 Cyber, which OpenAI describes as a 51% relative improvement.
Web penetration testing without source code: 40.3% recall of vulnerabilities against 32.0% for GPT-6 Astra.
RemoteExploitBench, OpenAI’s new test using known vulnerabilities in applications such as GitLab, PostgreSQL and Jenkins: GPT-6 Astra “refused 100% of tasks”, while GPT-6 Cyber “completed 95%”.
Staying in scope: on its ExploitGym Honeypot evaluation, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of cases without deployment safety prompts, up from 94.49% for GPT-5.6 Cyber, and 100% with them.
OpenAI says GPT-6 Cyber “outperforms GPT-6 Astra on most cyber tasks, even when Astra does not refuse the request.”
What Was Posted on X
Introducing GPTโ6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.
The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Ciscoโalongside our own at OpenAIโare already putting it to work.
GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time.
OpenAI posted this at 17:09 UTC on 5 November 2026 to introduce GPT-6 Cyber. It says the model powered the Linux kernel research OpenAI announced the day before, that security teams at companies including Mozilla, Palo Alto Networks, Calif and Cisco are already using it, and that it is available to select Daybreak Red customers, with access expanding over time.
HOLY: OpenAI is rolling out GPT-6 Cyber, and it BEATS flagship GPT-6 Astra on ExploitGym: 48.5% vs 42.4%.
It also helped find 81 previously unknown Firefox vulnerabilities. In safety testing, it went outside authorized targets less than 1% of the time, down from roughly 5% for GPT-5.6 Cyber.
Choblin, who calls it “HOLY” and says it “BEATS” GPT-6 Astra, posted this at 17:15 UTC on 5 November 2026. The post puts GPT-6 Cyber at 48.5% on ExploitGym against 42.4% for GPT-6 Astra, says it helped find 81 previously unknown Firefox vulnerabilities, and says that in safety testing it went outside authorized targets less than 1% of the time, down from roughly 5% for GPT-5.6 Cyber. OpenAI’s system card gives 48.45% and 42.43% on ExploitGym, and honeypot exploitation of 0.07% for GPT-6 Cyber against 5.51% for GPT-5.6 Cyber.
Scores on OpenAI’s own cyber evaluations, earlier model vs GPT-6 Cyber, as reported in OpenAI’s launch post on 5 November 2026. Each row uses the comparison modelOpenAI named for that test.
OpenAI says its researchers used GPT-6 Cyber to develop 948 privilege-escalation exploits in Linux, the run we covered in OpenAI’s AI Reproduced 1,724 Linux Kernel Flaws in $1.2M Run. The launch post says GPT-6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948, “more than five times the 166 exploited by GPT-5.6 Sol using the same research harness.” By our arithmetic, 948 is 5.7 times 166.
OpenAI also says the model “produced evidence and reproducible tests for 81 novel vulnerabilities classified by Mozilla as high severity” in Firefox, and that Mozilla began landing fixes. It names Cisco, Calif and CodeRabbit as security teams using the model for vulnerability research and code review.
Access and Safeguards
OpenAI splits access between two programs: Daybreak Blue, which it says “helps verified defenders validate vulnerabilities and fixes with fewer unnecessary refusals”, and Daybreak Red, which “enables deeper exploit development and authorized security testing with GPT-6 Cyber.”
IP allowlisting is required for Daybreak Red API access by default, OpenAI says.
OpenAI lists cyber-specific misalignment monitoring, targeted monitoring for emerging threats and Private Safety Processing among its safeguards.
A new Daybreak Red for Critical Infrastructure group is for approved defenders of systems such as power and water, and OpenAI says it worked with Dragos on industrial-control evaluations.
OpenAI points to its $1 billion Daybreak for Frontline Defenders grant program, which it says subsidizes access for resource-constrained teams.
Update, 6 Nov, 14:30 UTC: OpenAI says adversarial groups are trying to distill its models. The Wall Street Journal’s CIO Journal newsletter led its November 6 edition with GPT-6 Cyber under the headline “OpenAI Bolsters Cyber Defenses”. One line it highlighted from OpenAI’s launch post was not in our earlier coverage.
In that post, OpenAI wrote: “As models become more capable and uncover more security vulnerabilities, we are also seeing more attempts to use them to cause harm. These attempts include efforts by malicious cyber actors and adversarial groups seeking to distill our models.” OpenAI said it had “paired GPT-6 Cyber’s stronger capabilities with stronger safeguards to detect and disrupt misuse.”
The post does not name the groups or say how many distillation attempts it saw. The same post said OpenAI’s goal is “to dramatically expand Daybreak”, its cybersecurity program.
As we read the launch post, OpenAI is pairing the stronger model with narrower access: approved Daybreak Red customers first, with IP allowlisting required for API access by default.
As we read the benchmark list, the comparison model changes by test: GPT-5.6 Cyber on ExploitGym and the honeypot evaluation, GPT-6 Astra on web penetration testing and RemoteExploitBench.
As we read the post, OpenAI gives no price and no date for wider access.
OpenAI, 5 November 2026
“GPTโ6 Cyber is a major step up from GPTโ5.6 Cyber and outperforms GPTโ6 Astra on most cyber tasks, even when Astra does not refuse the request.”
Three Implications
FOR SECURITY TEAMS OpenAI names Cisco, Calif and CodeRabbit as users for vulnerability research and code review.
FOR OPEN-SOURCE MAINTAINERS OpenAI says Mozilla began landing fixes for Firefox vulnerabilities the model found.
FOR CRITICAL INFRASTRUCTURE OpenAI says a new Daybreak Red for Critical Infrastructure group is for approved defenders, and that it worked with Dragos on evaluations.
The framework says: “It is an operating system that wraps your AI โ a harness that compresses your thinking, your teamโs feedback, and the modelโs raw capability into something replicable.”
As we read it, OpenAI’s own comparison uses that idea: it credits GPT-6 Cyber with 948 exploits against 166 for GPT-5.6 Sol “using the same research harness”. The framework itself does not discuss GPT-6 Cyber.
Update, 5 Nov, 17:34Z: more announcements set for 12 November. In a post on X at 17:09 UTC, OpenAI said it will hold a “Cyber at the Frontier” livestream on 12 November at 10am PT “for more Cyber announcements” (event page).
Update, 5 Nov, 17:37Z: how GPT-6 Cyber compares with GPT-6 Astra. OpenAI’s GPT-6 Cyber system card puts its intended-exploit success rate on ExploitGym at 48.45%, compared with 42.43% for GPT-6 Astra, and at 36.94% against 29.69% on a hardened variant with stronger exploit mitigations. On the ExploitGym Honeypot test, without deployment safety instructions, GPT-6 Cyber exploited the out-of-scope honeypot in 0.07% of graded rollouts, compared with 5.51% for GPT-5.6 Cyber and 0% for GPT-6 Astra (system card).
What Is Not Established
All benchmark figures are OpenAI’s own, on evaluations OpenAI built or ran; we have not seen independent results. OpenAI does not give a price for GPT-6 Cyber or a date for wider Daybreak Red access. We have not read OpenAI’s system card in full.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
OpenAI’s GPT-6 Cyber scores 48.5% on the ExploitGym test against 32.1% for the previous cyber model, is rated Critical for cybersecurity under OpenAI’s framework, and is available at first only to approved Daybreak Red customers.
A note on sourcing. Figures and quotations are from OpenAI’s launch post of 5 November 2026. Benchmark results are OpenAI’s own; the 5.7x multiple is our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
