OpenAI Cyber Model Finds 81 Firefox Flaws, Builds 948 Linux Exploits

OpenAI has introduced GPT-6 Cyber, a restricted-access AI model that the company says found 81 previously unknown high-severity vulnerabilities in the Firefox browser and produced working exploits for 948 Linux kernel vulnerabilities. According to CSO Online, Mozilla has started releasing fixes, while the Linux findings are being sent to maintainers with proposed patches for review.

The results come from OpenAI's own researchers and have not been independently confirmed in the material CSO reviewed. Security experts quoted by the outlet say the findings are notable but argue that finding flaws is the easy part. Getting them fixed is harder.

The key numbers behind GPT-6 Cyber

• 81: Previously unknown high-severity Firefox vulnerabilities that OpenAI researchers say the model found. Mozilla has begun releasing fixes.
• 948: Linux kernel vulnerabilities for which the model helped develop working exploits, according to OpenAI.
• 588: The portion of those 948 Linux flaws that were previously unknown.
• 10 of 166: In a sample of 166 exploits, OpenAI's researchers identified 10 techniques that the company says appeared novel. External kernel exploitation experts reportedly had not seen them before. They could let someone with local access gain root privileges.
• 0.07% vs. 5.51%: How often the model strayed beyond its intended scope in OpenAI's ExploitGym Honeypot test, compared with the earlier GPT-5.6 Cyber model.
• Less than 25%: The remediation rate shown so far on OpenAI's Patch the Planet dashboard, according to Chris Hughes of Noma Security.

What OpenAI says the model is for

OpenAI describes GPT-6 Cyber as its most capable and best-aligned cybersecurity model. The company says it is meant to help security researchers find and validate exploitable flaws, assess their impact and test patches before release. It says the model was trained to stay within authorized limits, respect restrictions and report results accurately.

On a press call, OpenAI representatives said such models can help confirm whether a flaw is exploitable, look for other attack routes, strengthen patches and assist with deploying them. They described the goal as security working at "agent speed." Researcher Eric Wallace said one aim is to check exploitability at scale and rank the most consequential flaws. He added that a fix should be tested against other routes of attack, since one vulnerability might not be the only path in.

Clint Gibler, OpenAI's cybersecurity research lead, described advance coordination on vulnerabilities in PHP, a language widely used on websites. OpenAI worked with researchers, maintainers and security vendors. It also worked with Cloudflare so the company could deploy web application firewall rules that protect sites before the bug was announced.

Why it matters for people who use this software

Firefox and the Linux kernel run on many personal and business computers, and PHP underpins many websites. If an AI model can uncover serious flaws in code that has been reviewed for years, defenders may have more to fix than they expected. Hughes told CSO that, if Mozilla's account holds up, mature software hides more vulnerabilities than many assumed. He noted that the Firefox bugs surfaced after several generations of advanced models had already helped harden that code.

The concern is the gap between discovery and repair. CSO reported that OpenAI provided no measurements showing how much its models speed up patch deployment or reduce exposure for organizations. Hughes said Patch the Planet currently shows a remediation rate below 25%. He argued that better discovery could widen that gap, leaving more flaws found but unfixed. Those flaws could then be found by attackers with similar tools, or by those using clever setups with less capable models. He called for defender-focused benchmarks that measure progress on closing the gap, not just on finding bugs.

Rob T. Lee, the new chief AI officer at industrial cybersecurity firm Dragos, which recently partnered with OpenAI, pointed to the 10 novel techniques. He said they could indicate that models can create new exploitation paths, not just find more flaws, and that this would make detection more complex. CSO noted that Lee has just joined Dragos, whose founder and CEO is Rob M. Lee.

Who can use it, and what comes next

OpenAI classifies GPT-6 Cyber as critical for cybersecurity under its own Preparedness Framework. Access is limited. The model will first be offered through OpenAI's API and Codex to a select group of approved customers in Daybreak Red, the more permissive of two vetted tiers in the company's trusted-access cyber program. Daybreak Blue is the tier OpenAI recommends for most defenders.

OpenAI said it will expand GPT-6 Cyber access to all Daybreak Red customers over the coming weeks and is also making GPT-5.6 Cyber available to that group. It is broadening access to an alpha of Daybreak Blue for Astra, its general-purpose frontier model, as part of a push to bring advanced cyber tools to more organizations.

Hughes acknowledged the safety gain from the lower out-of-scope rate. But he cautioned that at the scale these systems operate, "even a small fraction means actions against systems nobody authorized."

Read next
Three Ex-OpenAI Safety Researchers Say Firings Were Retaliation; Company Disagrees
Google's Gemini 4 Carbon AI Targets Software Flaws, but Few Can Use It Yet
Hackers Leak Personal Details of Tens of Thousands of DHS Staff

Around Prague, based on reporting by CSO Online