{"article":{"slug":"openai-gpt-6-cyber-daybreak-red","kind":"model","source":{"title":"GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents","url":"https://openai.com/index/introducing-gpt-6-cyber/","date":"2026-11-05"},"headline":"GPT-6 Cyber scores 48.5% on ExploitGym, but only approved security teams can use it","whatChanged":"OpenAI announced GPT-6 Cyber on November 5. It is a cybersecurity variant of GPT-6 Astra for authorized security work. OpenAI says it is available through the API and Codex to a select group of approved Daybreak Red customers and will reach all of them \"soon\", with no date. Access needs identity verification, use restrictions, single sign-on, multi-factor authentication and API IP allowlisting. The model page lists $20 per million input tokens, $100 per million output tokens and $2 for cached input, with a 1,050,000-token context window.\n\nOpenAI reports 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, against 42.4% for GPT-6 Astra in the system card. On RemoteExploitBench, a new OpenAI-built test, Astra with Daybreak Blue access refused every task and GPT-6 Cyber completed 95%. OpenAI rates the model Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. The system card also reports an outside assessment. Irregular found performance broadly comparable to Astra's: 11 of 11 CyScenarioBench challenges solved at least once against 8 of 11, but 90 of 288 FrontierCyber challenges against 96, with no successful attacks on fully hardened targets.","whyItMatters":"The chart is mostly an access story. In OpenAI's defensive-use suite, standard GPT-6 Astra completes 2.4% of exploit-development tasks and 7.4% of red-teaming tasks, while the gated models complete 92% to 98% of exploit tasks. GPT-6 Cyber leads Astra by about 6 points on ExploitGym, while Irregular's outside test found the two broadly comparable.\n\nFew buyers can use the model directly. A company is more likely to meet it inside a penetration-testing or code-review vendor's service, so the practical question is which model that vendor runs and whose code it sends to OpenAI. OpenAI also opened the earlier GPT-5.6 Cyber to every Daybreak Red customer, so a security team that qualifies has a gated model today.","issuer":"OpenAI","glance":"OpenAI released GPT-6 Cyber November 5 to approved Daybreak Red customers only, at $20 in and $100 out per million tokens.","keyFact":"$20 in, $100 out; Daybreak Red only","action":"Ask your penetration-testing vendor which AI model it runs and under what authorization limits.","visuals":[{"url":"https://deploymentsafety.openai.com/data/eval-sets/gpt-6-cyber/assets/images/image2.png","alt":"Grouped bar chart of authorized cybersecurity task completion in four categories. In exploit development GPT-6 Cyber scores 96%, GPT-5.6 Cyber 98%, GPT-6 Astra with Daybreak Blue 92% and standard GPT-6 Astra 2.4%.","caption":"OpenAI's system card Figure 1, November 5. Task completion in OpenAI's own defensive-use suite, run by OpenAI, with no sample sizes shown. Exploit development: GPT-6 Cyber 96%, GPT-5.6 Cyber 98%, Astra with Daybreak Blue 92%, standard Astra 2.4%. Red-teaming: 97%, 100%, 76.9%, 7.4%. The gap follows access tier, and GPT-6 Cyber does not lead GPT-5.6 Cyber here.","credit":"OpenAI","source":{"title":"GPT-6 Cyber System Card","url":"https://deploymentsafety.openai.com/gpt-6-cyber","date":"2026-11-05"}}],"attachments":[{"title":"GPT-6 Cyber System Card","url":"https://deploymentsafety.openai.com/gpt-6-cyber","description":"OpenAI's safety documentation: the Critical cybersecurity designation, ExploitGym and related results, Private Safety Processing, access rules, and the NIST CAISSI and Irregular external evaluations, including Irregular's FrontierCyber result.","source":{"title":"GPT-6 Cyber System Card","url":"https://deploymentsafety.openai.com/gpt-6-cyber","date":"2026-11-05"}},{"title":"GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents","url":"https://openai.com/index/introducing-gpt-6-cyber/","description":"OpenAI's announcement: the access plan for Daybreak Red, the 48.5% ExploitGym result, RemoteExploitBench, the scope-alignment test and OpenAI's customer testimonials.","source":{"title":"GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents","url":"https://openai.com/index/introducing-gpt-6-cyber/","date":"2026-11-05"}},{"title":"GPT-6-Cyber model page, OpenAI API","url":"https://developers.openai.com/api/docs/models/gpt-6-cyber","description":"The API price list ($20 input, $100 output, $2 cached input per million tokens), the 1,050,000-token context window, the 128,000 output limit and the Daybreak Red approval requirement.","source":{"title":"GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents","url":"https://openai.com/index/introducing-gpt-6-cyber/","date":"2026-11-05"}}],"evidenceNote":"The ExploitGym chart in OpenAI's announcement is drawn inline on the page and has no image URL that could be verified, so the visual here is the system card's Figure 1 on defensive-use task completion. Its scores come from OpenAI, and public reuse terms were not checked beyond linking to the image.","companyLabel":"Public edition"},"issue":{"id":"2026-11-09","publishedLabel":"Week of November 9, 2026","isLatest":true},"position":{"index":1,"total":4},"next":{"slug":"anthropic-message-threads-beta","headline":"Anthropic's message threads cut first-token wait by about 35% on very long sessions"},"session":null}