OpenAI put its GPT-6 Cyber model at 52% (52.6% on its on-stage chart; clarified 13 November) on the ExploitGym benchmark during its Cyber at the Frontier keynote on 12 November, and named a rival figure: “For comparison, Anthropic’s Mythos 5.2, which released earlier this week, scores at about 38%”, an OpenAI speaker said at 21:19 in the video. That is 14 points, or about 1.4 times, by our arithmetic.
Update and correction, 13 November 2026, 13:00 UTC: OpenAI’s on-stage chart puts its unreleased next cyber model at 77.0% on ExploitGym. Frames of the keynote video at 21:48 and 21:54 show an “ExploitGym” slide plotting success rate against average output tokens: 77.0% for an “Upcoming Cyber Model”, 52.6% for GPT-6 Cyber and 42.4% for GPT-6 Astra. The model is unnamed and unreleased, and these are OpenAI’s own benchmark runs. This piece first gave GPT-6 Cyber at 52%, the figure spoken on stage; we have added the chart’s 52.6% to the opening paragraph. The slide shows no Anthropic or Googlemodel: the 39.2% for Mythos 5.2 and 54.8% for Gemini 4 Carbon in a widely shared post on X do not appear on it, and on stage OpenAI put Mythos 5.2 at “about 38%”. Our arithmetic: 77.0% is 24.4 points above 52.6% and about 2.0 times 38%, consistent with the “nearly double” claim.
The same speaker said OpenAI is “already working on the next specialized cyber model”, which on the same tasks scores “nearly double the score of Mythos 5.2”. The keynote gave that model no name, date or 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 does not discuss OpenAI’s or Anthropic’s cyber models.
The key insight: As we read the keynote, OpenAI changed the yardstick in a week: on 5 November it measured GPT-6 Cyber against its own GPT-5.6 Cyber at 48.5%, and on 12 November against Anthropic’s Mythos 5.2, at 52% to about 38%, with a next model it says nearly doubles Mythos 5.2.
ExploitGym “tests the model’s ability to create a proof of concept for a given vulnerability”, the speaker, introduced on stream as Fouad, said at 21:06. On GPT-6 Cyber: “We trained” it “to work harder and not give up as easily”, reaching 52%.
The next model is “even more capable, token efficient” and, like GPT-6 Cyber, “able to work harder and longer by default”, he said at 21:42. “With our upcoming cyber model, we’re continuing to reset the frontier of cyber capabilities beyond anything that is available today.”
The keynote’s opening speaker, introduced as Greg, had shown a chart of “our latest cyber model” at 13:28, a modelOpenAI is “working on releasing soon”. The stream did not say whether the two descriptions refer to the same model.
OpenAI’s Cyber at the Frontier keynote, 12 November 2026, on OpenAI’s YouTube channel. The benchmark passage runs from 21:02 to 22:35 in the video.
ExploitGym scores as OpenAI stated them: 52% for GPT-6 Cyber and about 38% for Anthropic’s Mythos 5.2 at the 12 November keynote, 48.5% and 32.1% in its 5 November launch post. The 14-point gap is our arithmetic.
Two OpenAI Numbers for One Test
The 52% is higher than the figure OpenAI published a week earlier. Its 5 November launch post put GPT-6 Cyber at 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber, as we reported then, and the system card gave 48.45%. The keynote did not say what changed between the two runs.
On stage, the 52% came straight after the line about training the model “to work harder and not give up as easily”. The comparison with Mythos 5.2 is OpenAI’s. Anthropic’s model has been scored on other tests: Cognition put it at 73.5 on its own internal cyber benchmark, which we covered on 10 November, a different test with a different scale.
OpenAI’s two ExploitGym comparisons a week apart: against GPT-5.6 Cyber in its 5 November launch post and against Anthropic’s Mythos 5.2 at the 12 November keynote. Point gaps are our arithmetic.
From Finding Bugs to Fixing Them
The second benchmark OpenAI showed was PatchBench, which “tests the model’s ability to find and fix a given flaw without causing a regression in the application”, the speaker said at 22:11. “while 6Astra performed reasonably well and 6Cyber had modest improvements”, the upcoming model scores “around 80% performance”, he said. The stream gave no PatchBench figure for GPT-6 Astra or GPT-6 Cyber.
OpenAI framed the shift in one line: “we believe the challenge is no longer just finding vulnerabilities. It’s making sure you can patch them as fast as you can find them.” The same keynote launched Defense Factory, which drafts patches for engineers, the product we covered earlier on 12 November.
Update (13 Nov): President Trump’s AI task force discussed AI-powered cyberattacks. Members of the task force “met on Thursday and discussed ways to encourage critical US sectors to report AI-powered cyberattacks”, according to a Trump administration spokesperson quoted by Reuters. Per the statement Reuters cites, vice chair Emil Michael, the Defense Department technology lead, “talked about progress on setting national standards for AI developers to report safety incidents and share information about cybersecurity”, and members discussed options for critical infrastructure providers to disclose vulnerabilities and attacks “without fear of legal consequences”.
Reuters adds that the 2015 Cybersecurity Information Sharing Act, which underpins threat-sharing between businesses and the government, “is set to expire next month”. No decision, order or deadline from the meeting was reported, and we found no official readout at our 03:1x UTC check; the details above rest on the statement as Reuters reports it.
Our read: incident reporting is the lever that touches the cyber models above. If labs must report safety incidents to a national standard, the defensive-cyber pitch of models like GPT-6 Cyber comes with a disclosure duty attached.
The Structural Read
As we read the PatchBench passage, OpenAI is moving the contest from finding flaws to fixing them, the job its new Defense Factory product sells, and it gave a number only for the model it has not released.
As we read the two ExploitGym figures, the 3.5-point difference between 48.5% and 52% is unexplained on stream, so we chart both and treat neither as the definitive score.
OpenAI speaker introduced as Fouad, Cyber at the Frontier keynote, 12 November 2026, 21:19
“For comparison, Anthropic’s Mythos 5.2, which released earlier this week, scores at about 38%.”
Three Implications
FOR SECURITY BUYERS OpenAI now quotes a rival’s score on stage; the test conditions behind the 38% were not given, so a like-for-like comparison needs Anthropic’s own number.
FOR ANTHROPIC OpenAI used Mythos 5.2, which the speaker said released earlier this week, as its reference model, the same model Cognition put at 73.5 on its internal benchmark.
FOR OUR COVERAGE We will add the next cyber model’s name, date, price and exact scores if OpenAI publishes them.
The framework describes “The Verifier Moat” as one “built by RLVR on domain-specific reward signals, building into a new asset class for reasoning-heavy verticals.”
As we read it, an exploit either works or it does not, which makes cyber a domain with a clear reward signal, and benchmarks such as ExploitGym are where labs now compete in public. The framework predates the keynote and does not discuss it.
What Is Not Established
The 38% for Mythos 5.2 is OpenAI’s figure, given on stage without the test conditions; we have not seen Anthropic publish an ExploitGym score for it. OpenAI gave no exact ExploitGym score for its next model, only “nearly double”; doubling 38% would be 76%, our arithmetic. Posts on X speculated about a new GPT model around the stream; the keynote did not name one. We transcribed the keynote ourselves because the video has no captions; quotes follow our transcript.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
OpenAI used its cyber keynote to score itself against Anthropic: 52% for GPT-6 Cyber on ExploitGym against about 38% for Mythos 5.2, by OpenAI’s numbers, with an unnamed next model at nearly double Mythos 5.2 and around 80% on PatchBench. No date, name or price came with it.
A note on sourcing. OpenAI’s keynote video has no captions, so we downloaded its audio and transcribed it ourselves on 12 November 2026; quotes and timestamps follow that transcript, with the benchmark figures re-checked against a second, larger transcription model. The 5 November figures come from OpenAI’s GPT-6 Cyber launch post as we reported it; the Cognition figure from Cognition’s post. Point gaps and ratios are ours.
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.
