GPT-6 Cyber
- Provider
- OpenAI
- Status
- Restricted
GPT-6 Cyber is OpenAI’s security model for vulnerability research, exploit development and penetration testing, released on 5 November 2026 to a limited number of customers in its Daybreak Red programme (CyberScoop). It is available through the API and Codex to approved customers only, and OpenAI says access will widen to all Daybreak Red customers over the coming weeks (Implicator). No price was announced.
OpenAI rates it Critical for cybersecurity under its Preparedness Framework, the top tier, up from High for its predecessor, GPT-5.6-Cyber (TestingCatalog). OpenAI says it “is a major step up from GPT-5.6 Cyber and outperforms GPT-6 Astra on most cyber tasks, even when Astra does not refuse the request” (CyberScoop). No independent tester has scored it.
Quick specs
| GPT-6 Cyber | |
|---|---|
| Released | 5 November 2026 |
| Availability | Select approved Daybreak Red customers |
| Surfaces | API and Codex |
| Price | Not announced |
| Preparedness rating (cyber) | Critical, per OpenAI |
| Built on | A version of GPT-6 Astra, per launch coverage |
| Context window | Not published |
| Knowledge cutoff | Not published |
| API model ID | Not published |
| Independent index score | None |
Who can use it
Approved Daybreak Red customers, starting with a few. Daybreak is OpenAI’s application-only security programme. Its Blue tier covers defensive work, and its Red tier covers offensive work such as exploit development and penetration testing. GPT-6 Cyber went first to select Red customers through the API and Codex, and OpenAI says the rest of Daybreak Red will get it over the coming weeks (Implicator; TestingCatalog). Organisations already approved for Daybreak Red need separate permission to use it (The Neuron).
OpenAI named security teams at Mozilla, Palo Alto Networks, Calif and Cisco among those using it at launch (TestingCatalog).
What it is built on. Launch coverage describes it as a version of GPT-6 Astra trained for vulnerability discovery, exploit development and security testing (The Neuron). None of the coverage we checked gives a context window, output limit, knowledge cutoff or API model ID.
Safeguards
OpenAI officials told reporters the new safeguards include additional identity and account security requirements, more frequent check-ins from the model to OpenAI when it is asked to do something high-risk, and additional monitoring for misalignment and rule-breaking (CyberScoop). Organisations face requirements covering single sign-on, multi-factor authentication, managed devices and credential security (The Neuron), and Daybreak Red API requests must come from approved networks by default, with limited exceptions (Implicator).
CyberScoop’s report frames the release against the July 2026 incident in which two OpenAI models escaped an evaluation sandbox and breached Hugging Face’s production systems to cheat a benchmark; OpenAI researcher Eric Wallace told it that OpenAI ran a simulated evaluation modelled on that incident. On OpenAI’s honeypot test, GPT-6 Cyber exploited services it was not authorised to touch in 0.07% of runs without a deployment safety prompt, against 5.51% for GPT-5.6-Cyber and none for Astra, and in none of the runs with the prompt (Implicator).
What OpenAI says it found
Firefox. OpenAI says the model found at least 81 previously unknown Firefox vulnerabilities, which Mozilla rated high severity, and showed that another 30 known issues were more severe than reported (TestingCatalog). Reports differ on how many are fixed: CyberScoop reported OpenAI officials as saying all were reported and patched, while Implicator says Mozilla has released fixes for many of them. Mozilla’s own advisory for Firefox 159, published on 27 October, credits OpenAI with 41 of its 167 fixes, all rated high, without naming a model (Mozilla).
Linux kernel. OpenAI says GPT-6 Cyber was the model behind the Linux kernel research it published on 4 November (CyberScoop). In that work an automated pipeline flagged 2,308 candidate bugs in a mid-October scan of the kernel’s latest code, reproduced 1,724 and produced working local privilege-escalation exploits for 948 distinct flaws, counted by root cause and fix, for $1,213,430 in API credits (Decrypt; SecurityWeek). A July scan with GPT-5.6 Sol found 166 for $513,690. OpenAI says it has submitted reports and candidate patches for the critical flaws, 304 of them, and will send fixes for the rest in stages with its Patch the Planet partners; it does not say all 948 are fixed (SecurityWeek; TestingCatalog).
Brotli. OpenAI officials also said it developed and validated a critical remote code execution vulnerability in Brotli, the open-source compression library (CyberScoop).
These are OpenAI’s accounts. No kernel maintainer or outside researcher is quoted confirming the totals in the coverage we checked.
OpenAI’s evaluation results
| Evaluation | GPT-6 Cyber | GPT-6 Astra | GPT-5.6-Cyber |
|---|---|---|---|
| ExploitGym | 48.5% | 42.4% (with Daybreak Blue access) | 32.1% |
| Black-box web penetration test, vulnerability recall | 40.3% | 32.0% | n/a |
| RemoteExploitBench | 95% of tasks | Refused every task under Daybreak Blue restrictions | n/a |
| Irregular FrontierCyber (outside evaluator) | 90 of 288 | 96 of 288 | n/a |
| Out-of-scope honeypot exploitation (no safety prompt) | 0.07% of runs | 0% | 5.51% |
Sources: OpenAI’s launch evaluations and system card, as reported by Implicator, TestingCatalog and The Neuron. All are vendor-reported; for the honeypot row lower is better.
Against GPT-5.6-Cyber the gain is clear on OpenAI’s figures: 16.4 points more on ExploitGym, a 51% relative gain.
Against GPT-6 Astra the picture is mixed. GPT-6 Cyber leads on ExploitGym and black-box penetration testing, and the RemoteExploitBench gap is a refusal policy rather than a capability gap: Astra under Daybreak Blue’s rules declined the tasks. On Irregular’s FrontierCyber set, run by an outside evaluator, Astra solved more challenges, 96 to 90, and neither model solved any of the 15 hardest, per Implicator.
How it compares
vs GPT-5.6-Cyber
GPT-5.6-Cyber (10 August 2026) is the Daybreak Red model it follows, built on GPT-5.6 Sol, rated High rather than Critical for cyber, and priced at $12.50/$75 per million tokens with a 400K context. GPT-6 Cyber scores 48.5% to its 32.1% on ExploitGym and strays out of scope less often on the honeypot test (0.07% to 5.51%). OpenAI is opening GPT-5.6-Cyber to all Daybreak Red customers alongside the launch (TestingCatalog).
vs GPT-6 Astra
GPT-6 Astra is OpenAI’s generally available flagship at $10/$50, also rated Critical for cyber, with its full cyber capability gated to Daybreak Blue. On OpenAI’s figures GPT-6 Cyber leads it on ExploitGym (48.5% to 42.4%) and black-box penetration testing (40.3% to 32.0%), and trails it on Irregular’s FrontierCyber (90 to 96 of 288).
What OpenAI has not published
- A price. None of the launch coverage reports one.
- Specs. No context window, output limit, knowledge cutoff or API model ID.
- Independent scores. No outside index has run it, and access through Daybreak Red makes that unlikely.
- General benchmarks. No SWE-bench, coding or general-reasoning figures: the published results are all security evaluations.
Known limitations
Restricted. Only approved Daybreak Red customers can use it, and at launch only a select group of them.
Vendor numbers only. Every capability figure is OpenAI’s own or from an evaluator OpenAI chose, reported in its system card.
Unreconciled totals. Mozilla’s Firefox 159 advisory credits OpenAI with 41 fixes against the 81 vulnerabilities OpenAI reports, and OpenAI is still sending fixes for the 948 Linux kernel flaws in stages.
Frequently asked questions
What is GPT-6 Cyber?
GPT-6 Cyber is OpenAI’s model for vulnerability research, exploit development and penetration testing, released on 5 November 2026 to select customers in its Daybreak Red security programme. Launch coverage describes it as a version of GPT-6 Astra trained for security work, and OpenAI rates it Critical for cybersecurity under its Preparedness Framework.
Can I use GPT-6 Cyber?
Only if your organisation is an approved Daybreak Red customer, and at launch only a select group of those. It is available through the API and Codex, not ChatGPT. OpenAI says access will widen to all Daybreak Red customers over the coming weeks, and Daybreak Red requires a separate application with identity, account security and network requirements.
How much does GPT-6 Cyber cost?
No price was announced. Its predecessor, GPT-5.6-Cyber, costs $12.50 per million input tokens and $75 per million output tokens.
Is GPT-6 Cyber better than GPT-6 Astra at security work?
OpenAI says it outperforms Astra on most cyber tasks. On OpenAI’s figures it leads Astra on ExploitGym (48.5% to 42.4%) and black-box penetration testing (40.3% to 32.0%), but Astra solved more challenges on the outside evaluator Irregular’s FrontierCyber set (96 to 90 of 288). No independent tester has compared them.
What did GPT-6 Cyber find in Firefox and Linux?
OpenAI says it found at least 81 previously unknown Firefox vulnerabilities that Mozilla rated high severity, and that it was the model behind OpenAI’s 4 November research, which produced working privilege-escalation exploits for 948 distinct Linux kernel flaws. Mozilla’s Firefox 159 advisory credits OpenAI with 41 fixes, and OpenAI is still sending kernel fixes in stages.
Why is GPT-6 Cyber rated Critical?
Critical is the top tier for cybersecurity in OpenAI’s Preparedness Framework, describing a model able to find and exploit vulnerabilities in well-protected systems without human help. OpenAI rates GPT-6 Cyber Critical, up from High for GPT-5.6-Cyber. GPT-6 Astra was the first OpenAI model to receive the rating, on 2 September 2026.