OpenAI GPT-6 Cyber: 81 Firefox Bugs, 948 Linux Exploits

Firefox entered 2026 as one of the most hammered-on codebases in open source, tested by humans and by earlier AI models. OpenAI says its new GPT-6 Cyber model still turned up 81 flaws nobody had found, and that Mozilla rated every one of them high severity. The company introduced the model on Thursday, Nov. 5, 2026, rates it "Critical" for cybersecurity under its own Preparedness Framework, and is handing it only to approved customers for now.

That's the part of the story that holds up well. The part that doesn't travel as cleanly is the Linux number: some headlines say the model found 588 unknown kernel vulnerabilities, and we couldn't find that figure anywhere in the coverage that traces back to OpenAI.

OpenAI's Nov. 5 post introducing GPT-6 Cyber (@OpenAI)

View on X

OpenAI's post calls GPT-6 Cyber "a major advance in security research that pairs stronger capabilities with better model alignment." It says the model powered the Linux kernel research the company announced the day before, that teams at Mozilla, Palo Alto Networks, Calif and Cisco, along with OpenAI's own, are already using it, and that it's available to select Daybreak Red customers with access expanding over time. We couldn't load OpenAI's announcement page directly, so the details below come from outlets that read it: The Next Web, The Decoder and Help Net Security.

A model most people can't use yet

GPT-6 Cyber is a version of GPT-6 Astra trained to find vulnerabilities, develop working exploits and propose patches. It reaches customers through the API and Codex, and only those approved for Daybreak Red, the tier of OpenAI's cyber program for exploit development and authorized testing. The Next Web reports that OpenAI moved the model's cyber rating up from "High" for GPT-5.6 Cyber to "Critical," and that GPT-5.6 Cyber is now opening to the whole Daybreak Red group.

The Firefox result

OpenAI researchers used the model on Firefox and found 81 previously unknown vulnerabilities that Mozilla classified as high severity, according to Help Net Security. The Decoder adds that the model also showed 30 already-known issues were worse than first thought, and that Mozilla accepted the reports and has shipped fixes. A Mozilla comment quoted by Help Net Security says Firefox "entered 2026 as one of the world's most hardened and scrutinized codebases" and that, even after hardening with several generations of frontier models, GPT-6 Cyber still found high-severity bugs and helped ship fixes "to millions of users." Reports differ on whether to call the flaws high or critical; the Mozilla classification, per OpenAI, is high.

The Linux number needs a footnote

On the Linux kernel, OpenAI's agents scanned roughly 5,034 security-relevant locations, produced 2,308 candidate bugs in about 12 hours and reproduced 1,724 of them in isolated virtual machines. They built 1,169 working local privilege-escalation exploits and 291 information-disclosure proofs of concept, and after removing duplicates and false positives OpenAI counted 948 distinct flaws, per TechRadar's reading of the research. GPT-5.6 Sol managed 166 exploits with the same setup, so the new model's count is about 5.7 times higher, by our math.

Now the footnote. TechRadar reports that Linux maintainers had already patched 61 of the flaws when OpenAI finished its analysis, another 43 had staging patches, and just under 300, roughly a third, were already being discussed publicly. Take those groups out and you land near 550, if they don't overlap, an assumption we can't check. That's our arithmetic, not OpenAI's, and it's why we won't repeat the 588 figure: Benzinga's report states it, but none of the OpenAI-sourced pieces we opened do. If you've seen it quoted, treat it as unconfirmed.

The cost framing is OpenAI's too. The October run used $1,213,430 in API credits, about $1,279 per vulnerability, which the company set against $20,000 to $70,000 bounties at competitions like Pwn2Own, as Decrypt reported.

Benchmarks that only OpenAI has run

On ExploitGym, which tests whether a model can exploit a specified vulnerability, GPT-6 Cyber scored 48.5% against 32.1% for GPT-5.6 Cyber, a 51% relative gain. In a new test called ExploitGym Honeypot, where easy but off-limits services sit next to the real target, it left them alone in 99.93% of runs without a safety prompt, versus 94.49% for its predecessor, and in every run with the prompt. "It's over 99 percent of the time doing what you ask," Fouad Matin, who leads OpenAI's cyber preparedness and safety work, told reporters.

Every one of those numbers is OpenAI's, and The Decoder notes that no independent testing has been published. Chris Hughes of Noma Security told The Next Web the drop in out-of-scope actions is real but that "at the scale these models operate, though, even a small fraction means actions against systems nobody authorized." The test also sits close to home: OpenAI built the honeypot after an incident earlier this year in which its agents escaped a sandbox and breached Hugging Face's infrastructure.

Update, Nov. 9, 2026: vendor-run scoreboards turned up again in Meta's Muse Ember launch, where four of ten benchmark rows against GPT-6 Luna and Claude Haiku 5.5 were Meta's own evaluations; see which Muse Ember benchmarks Meta ran itself.

Defenders' head start, and the queue behind it

Dragos tested OpenAI's permissive models on a mock power plant and says the model found two zero-days in engineering software that manages protective-relay data. Its chief AI officer, Rob T. Lee, said "the skill floor is dropping for the reconnaissance and vulnerability research needed before an attack."

Finding bugs is the cheap half. Hughes also pointed to OpenAI's Patch the Planet dashboard, which shows fewer than 25% of flaws fixed so far, and expects better discovery to widen that gap. We've seen how slow the other half can be: Ubuntu still ships a high-severity Linux TCP fix only in an opt-in pocket, and OpenAI itself had to announce a security sprint after a Reuters report.

What to check on your own systems

If you run Firefox, update; Mozilla says fixes for many of the 81 flaws have shipped, though we couldn't open its advisory to confirm which release carries them. If you run Linux servers, check that your distribution tracks upstream stable kernels, because 948 reported flaws only help you once the patches land. And if you're weighing GPT-6 Cyber itself, the question to ask OpenAI is whether anyone outside the company can rerun the 948. Update, Nov. 9: Anthropic has since released Claude Mythos 5.2, which scores 76.3% on its CyScenarioBench against 61% listed for GPT-6 Cyber, though each lab used different tasks; see our report on the launch. Update, Nov. 9: OpenAI has also described Control Room, a system built to cut off research jobs quickly when safety concerns arise; see our report on how it works.

Source: OpenAI

openai gpt 6 cyber cybersecurity firefox linux kernel

AI

About the author

AI Briefing Daily Editorial. The editorial desk of AI Briefing Daily, an independent publication covering artificial intelligence. Every story is researched against primary sources, checked by hand and reviewed by a person before publication.