Security
GPT-6 Cyber in numbers: rated Critical, 588 of 948 Linux exploits target unknown flaws, 304 critical findings sent to maintainers with candidate patches
OpenAI rates GPT-6 Cyber Critical. Of 948 Linux exploits, 588 hit unknown flaws; 304 critical findings went to maintainers with candidate patches, per OpenAI.
- Models
- #OpenAI
- #GPT-6 Cyber
- #Linux kernel
- #Preparedness Framework
- #Daybreak Red
- #vulnerabilities
- #cybersecurity
OpenAI rates its GPT-6 Cyber model Critical, the highest tier of its Preparedness Framework for cyber capabilities. According to OpenAI, as relayed by the press, a research run reproduced 1,724 Linux kernel vulnerabilities and built working exploits for 948 of them, 588 of which targeted previously unknown bugs, and 304 critical findings have been disclosed to maintainers with candidate patches. All of these figures come from OpenAI and none has been independently verified. This piece decodes the numbers; it follows our earlier articles on the model's release and on the $1.2 million Linux pipeline.
The essentials
- Critical: the highest cyber-risk tier in OpenAI's Preparedness Framework; GPT-5.6 Cyber was rated High.
- 948 and 588: working exploits for 948 Linux kernel vulnerabilities out of 1,724 reproduced; 588 of the 948 (about 62%) targeted previously unknown bugs, which leaves 360 that were not new.
- 304: critical findings disclosed to Linux maintainers with candidate patches. "Candidate" means proposed for review, not merged.
- Five times more: 948 exploits against 166 for GPT-5.6 Sol under the same set-up, per OpenAI.
- Access: only select Daybreak Red customers, through the API and Codex, for now.
What does the Critical rating mean?
Critical is the highest risk tier for cyber capabilities in OpenAI's Preparedness Framework. According to The Register, it means a model can find and develop working zero-day exploits in real-world critical systems and carry out end-to-end cyberattacks. GPT-6 Cyber is rated Critical, up from High for GPT-5.6 Cyber, and The Decoder notes it is the same rating as its base model, GPT-6 Astra. The rating is OpenAI's own classification of capability, not a finding that an attack has taken place, and it comes with restricted access: GPT-6 Cyber is available through the API and Codex to a select group of approved Daybreak Red customers, with wider access for Daybreak Red customers planned over the coming weeks.
How do 1,724, 948, 588 and 304 fit together?
Per OpenAI as relayed: researchers reproduced 1,724 Linux kernel vulnerabilities and built working exploits for 948 of them (about 55%), more than five times the 166 that GPT-5.6 Sol managed with the same set-up. Of those 948, 588 (about 62%) were previously unknown bugs, which leaves 360 that were not new; the coverage we read does not describe those 360. Separately, OpenAI says it has disclosed 304 critical findings to Linux maintainers. The coverage does not say whether the 304 are part of the 588, only that they are critical and came with candidate patches. Plausibly they are the most severe reports sent first, as OpenAI said of its earlier pipeline, but that is our inference. The counts also come from different moments: OpenAI earlier said that ten days after the analysis 61 local privilege escalation flaws were fixed in the mainline kernel and 299 had been reported and publicly discussed, and we do not add those to the 304.
What does "sent with fixes" mean?
The Register says the Linux findings are being reported with potential patches for maintainer review. A candidate patch is a proposal: the kernel maintainers decide whether it is correct and whether to merge it, and a fix only protects users once it is merged and shipped in the kernels they run. For the 81 high-severity Firefox flaws, Mozilla has started fixing them (The Register). So "304 critical flaws transmitted with fixes" should be read as 304 critical findings reported with proposed patches, not as 304 flaws fixed.
What is not established
- Independent verification of the 948, 588 and 304 figures: all come from OpenAI through press relays.
- Whether the 304 overlap with the 588, and how many patches have already been merged upstream.
- What "previously unknown" is measured against (which database, as of which date), and what the other 360 were.
- Whether several of the 948 exploits target the same underlying flaw: the grouping into unique flaws is OpenAI's, and we could not check it.
- How far the results carry over to software other than the kernel version and virtual-machine set-up that OpenAI tested.
Summary table
| Point | What is stated as of November 6, 2026 | Source |
|---|---|---|
| Rating | Critical (GPT-5.6 Cyber: High) | The Register, The Next Web |
| Reproduced and exploited | 1,724 reproduced, 948 working exploits | The Next Web, The Decoder |
| Previously unknown | 588 of the 948 | The Register |
| Critical findings disclosed | 304, with candidate patches | The Next Web, The Decoder, Gulf Business |
| Comparison | GPT-5.6 Sol: 166 exploits with the same set-up | The Decoder |
| Access | Select Daybreak Red customers, API and Codex | The Register |
FAQ
Does Critical mean the model has been used in attacks?
No. It is OpenAI's classification of capability, and access is limited to vetted Daybreak Red customers. No misuse is reported in the coverage we read.
Are the 304 critical flaws fixed?
Not established. They were reported with candidate patches, which maintainers still have to review and merge.
Does 588 of 948 mean most flaws found were new?
It means about 62% of the exploited kernel bugs were previously unknown. The other 360 were not new, and we do not know what they were.
Method and limits
Sources checked on November 6, 2026. We read The Register through an automated extract (it contains the Critical rating, the 948 and the 588, but not the 304 or the 1,724) and search extracts of The Next Web, The Decoder, Gulf Business and a developer guide; OpenAI's own report and system card were not read. All figures are OpenAI's. This is not security advice.
Sources
- The Register: OpenAI releases GPT-6 Cyber, and promises this one stays within rules of engagement (verified 2026-11-06)
- The Next Web: OpenAI's GPT-6 Cyber found 81 new Firefox flaws and is rated 'Critical' (verified 2026-11-06)
- The Decoder: OpenAIs GPT-6 Cyber can both break into systems and propose how to fix them (verified 2026-11-06)
- Gulf Business: OpenAI launches GPT-6 Cyber with 'Critical' cybersecurity rating, uncovers 81 Firefox vulnerabilities (verified 2026-11-06)
- Byteiota: GPT-6 Cyber Found 948 Linux Exploits: Developer Guide (verified 2026-11-06)
Spotted an error in this article? Report it · Our corrections policy
