GPT-6 Cyber: On November 5, 2026, OpenAI introduced a specialized model for authorized vulnerability discovery, exploit development, and web penetration testing. Its safeguards are intended to keep security work within the user’s authorized scope, while initial access remains restricted.
How GPT-6 Cyber fits into authorized defense
AI-assisted vulnerability research is useful only when paired with permission, verification and responsible disclosure.
What Happened
OpenAI’s GPT-6 Cyber security model is a specialized variant of GPT-6 Astra for authorized security research. OpenAI reported an ExploitGym increase from 32.1% for GPT-5.6 Cyber to 48.5% for GPT-6 Cyber: 16.4 percentage points, or about 51% relative to the earlier score. The company also described research that identified 81 previously unknown high-severity Firefox vulnerabilities and, after reproducing 1,724 Linux kernel vulnerabilities, developed working exploits for 948 of them in a controlled harness. These figures describe OpenAI’s research and evaluations, rather than measurements independently reproduced by DigiOffly.
Who Is Affected & When
GPT-6 Cyber availability is initially limited to a select group of approved Daybreak Red customers through the API and Codex.
What Changes in Practice
Responsible use in security workflows
For defenders, a responsible workflow starts with an owned repository, isolated environment, or explicitly authorized assessment. Qualified reviewers should reproduce findings, coordinate disclosure with maintainers, and confirm that proposed patches solve the underlying issue.
Benchmarks and ExploitGym results
GPT-6 Cyber benchmarks include ExploitGym, which measures exploitation of specified vulnerabilities in controlled targets. This result is not a universal cybersecurity performance score. Separately, independent evaluator Irregular reported that GPT-6 Cyber solved 90 of 288 FrontierCyber challenges, compared with 96 of 288 for GPT-6 Astra.
Firefox vulnerabilities and Linux exploit research
The Firefox figure refers to 81 novel vulnerabilities classified by Mozilla as high severity; OpenAI disclosed the findings and proposed patches to Mozilla, which began landing fixes shortly after receiving the reports. The Linux figure refers to 948 working privilege escalation exploits developed from 1,724 vulnerabilities reproduced in a controlled research harness; it does not mean that 948 previously unknown vulnerabilities were simultaneously exposed across the internet.
Access and pricing
The cited OpenAI announcement and system card do not list public pricing. Organizations considering approved Daybreak Red use should confirm eligibility, commercial terms, and deployment requirements directly with OpenAI.
Safety and scope limitations
OpenAI says it trained the model to respect authorized scope and deployed access restrictions and monitoring. It also recommends sandboxing, action review, and human oversight to limit misuse. During ExploitGym Honeypot testing, the model avoided exploiting out-of-scope services in 99.93% of runs without deployment safety instructions and 100% with them. Those results do not guarantee the same behavior in every deployment.
| Measure | GPT-6 Cyber | Context / limitation |
|---|---|---|
| ExploitGym | 48.5% | GPT-5.6 Cyber: 32.1%; vendor-reported |
| Web pentest vulnerability recall | 40.3% | GPT-6 Astra: 32.0%; separate benchmark |
| FrontierCyber (Irregular) | 90 / 288 | GPT-6 Astra: 96 / 288; independent evaluator |
| Firefox research | 81 vulnerabilities | Disclosure and fixes described by OpenAI |
| Linux research | 948 working privilege escalation exploits | 1,724 vulnerabilities reproduced in a controlled harness |
Why It Matters
GPT-6 Cyber could help security teams reduce the time between identifying a vulnerability, validating its impact, and developing or testing a patch. However, benchmark scores differ by task and do not prove broad superiority. Restricted access, scope controls, and monitoring reflect the dual-use nature of exploit development. Security teams should evaluate these capabilities within a broader program of asset inventory, patch management, and independent risk review.
Our Take
The results indicate a meaningful advance on several research tasks, but Irregular’s separate FrontierCyber assessment illustrates why no single benchmark can establish superiority for every security workflow. Teams should assess legal authorization, access controls, reproducibility, and patch validation before adopting such capabilities.
Frequently Asked Questions
Is GPT-6 Cyber available in ChatGPT?
OpenAI has not announced general ChatGPT availability. Access is initially restricted to eligible Daybreak Red customers through API and Codex.
What is GPT-6 Cyber pricing?
No public price list appears in the cited materials. Organizations seeking access must first meet Daybreak Red eligibility and the additional approval required for GPT-6 Cyber.
What did OpenAI report about Firefox vulnerabilities?
OpenAI reported 81 novel Firefox vulnerabilities classified as high severity by Mozilla and supplied reproducible tests and proposed patches. Mozilla began fixing the issues after disclosure.
Have the safeguards been independently tested?
OpenAI’s system card documents its internal scope-compliance and safeguard testing. Irregular separately evaluated cybersecurity capabilities—not the safeguards—across CyScenarioBench, FrontierCyber, and Atomic Challenges. No independent safeguard audit is established by these cited evaluations.
Official Sources & Dates
Announced and initially made available to selected approved Daybreak Red customers on November 5, 2026. Check official updates before purchasing or changing production settings.
Editorial note: This analysis draws on OpenAI’s November 5, 2026 announcement and system card. Capability and safeguard results are vendor-reported unless an external evaluator is explicitly identified; DigiOffly did not conduct hands-on model testing. Prices and access conditions may change.