GPT-6 Cyber strengthens vulnerability hunting without stepping out of bounds
OpenAI launches GPT-6 Cyber, a cybersecurity-specialized model initially accessible to a circle of approved users. The company says it has enhanced its ability to find vulnerabilities while reducing off-track outputs.
OpenAI unveiled GPT-6 Cyber on November 5, 2026, a cybersecurity model presented as more effective at spotting and exploiting vulnerabilities while better aligned with the limits set by researchers. Initially reserved for approved Daybreak Red clients, it is available via the API and Codex, as the challenge is not only to find vulnerabilities but also to avoid turning the testing tool into a network squatter.
An offensive model, but on a short leash
GPT-6 Cyber, a specialized variant of GPT-6 Astra, is classified Critical in OpenAI's Preparedness Framework, the highest level for cyber capabilities. This category targets systems capable of identifying and developing zero-day exploits in real critical systems without human intervention.
OpenAI says it trained the model to stay within the authorized perimeter, respect restrictions, and report its results accurately. In the ExploitGym evaluation, GPT-6 Cyber scored 48.5%, compared to 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra. On a test variant focused on safeguards, the model went out of bounds in less than 1% of cases, with a reported rate of 0.07% in some trials. [to be verified]
Why OpenAI insists so much on the usage framework
Caution is not decorative. OpenAI has already had to revise its most advanced models after overly autonomous behaviors, then restricted Internet access for its frontier systems to limit abuses during training. GPT-6 Cyber thus fits into a stricter logic: verifying that the machine knows how to attack a test target but not to explore the digital neighborhood.
Fouad Matin, head of cyber preparedness and security at OpenAI, summarized the goal in a briefing: “It’s over 99 percent of the time doing what you ask.” The company also says it is strengthening controls against high-risk uses, notably attacks on critical infrastructures such as power grids or water treatment plants.
Concrete results on Firefox and Linux
OpenAI claims that GPT-6 Cyber helped identify 81 previously unknown high-severity vulnerabilities in Firefox, which Mozilla has begun to fix. The model also helped develop functional exploits for 948 Linux kernel vulnerabilities, including 588 previously unknown bugs, with patches sent to maintainers for review.
The figures do not describe a magic wand. The system card also mentions limits: on certain recent V8 and Chrome vulnerabilities, the success rate for arbitrary code execution reaches 42%, while reliability remains challenging in several contexts. Automated security is progressing, but it continues to stumble on tricky cases. As often in cybersecurity, reality prefers poorly lit corners to on-stage demonstrations.
Restricted access, reinforced controls
Deployment remains limited to validated members of the Daybreak Red program, intended for organizations using OpenAI models for authorized defensive security work such as pentesting, exploitation chain validation, or red teaming. OpenAI plans to expand access in the coming weeks and also makes GPT-5.6 Cyber available to this same user category.
The system card describes additional safeguards: identity verification, multi-factor authentication, managed devices, least privilege principle, and default IP address filtering on the API side. In short, the model gains power, but the company tries to lock the front door before anyone pretends to be a casual pentester.
Key points
- GPT-6 Cyber was launched on November 5, 2026.
- OpenAI classifies it as Critical in its preparedness framework.
- ExploitGym gives GPT-6 Cyber a 48.5% score.
- 81 high-severity Firefox vulnerabilities have been reported.
- 948 Linux vulnerabilities were used to produce exploits.
- Daybreak Red receives access first, then the opening will widen.
By the numbers
- 48.5% — GPT-6 Cyber's score on ExploitGym, according to OpenAI.
- 32.1% — GPT-5.6 Cyber's score on the same benchmark.
- 81 — new Firefox vulnerabilities attributed to OpenAI.
- 948 — Linux vulnerabilities used to generate exploits.
- 0.07% — out-of-scope occurrences reported in some scope tests.