OpenAI GPT-6 Cyber Finds 81 High-Severity Firefox Flaws and Shows Why Patch Backlogs Are About to Get Worse
OpenAI released GPT-6 Cyber on November 5, its fourth security model this year and the first it rates “Critical” for cybersecurity under its own Preparedness Framework. Researchers using it found 81 previously unknown high-severity flaws in Firefox. Mozilla has started shipping fixes. Against the Linux kernel, the model reproduced 1,724 known vulnerabilities and built working exploits for 948 of them. The previous model, GPT-5.6 Sol, managed 166 in the same test.
Access is limited for now. Only approved customers in Daybreak Red, OpenAI’s tier for exploit development and authorized security testing, can use it through the API and Codex. OpenAI says it will open it to all Daybreak Red customers soon. Dragos, the industrial security firm, tested it against a mock power plant and found two zero-days in engineering software that manages protective relay data.
For an IT team, the number that matters most is how fast those findings turn into work.
Finding is getting cheap, fixing isn’t
Vulnerability management has always been a race between discovery and remediation, and discovery has been the slow side. A good researcher might find a handful of serious bugs in a mature codebase in a year. A model that finds 81 in a browser that has been fuzzed and audited for two decades changes the ratio.
The fixes still go through people. A vendor has to confirm the bug, write a patch, test it and ship it. Then every company running that software has to schedule the update, test it against its own systems, and roll it out. Chris Hughes of Noma Security points to Patch the Planet data showing fewer than a quarter of the flaws found so far have been fixed, and he expects better discovery to widen that gap. OpenAI says it has sent 304 critical Linux findings to maintainers with proposed patches. Proposed patches help. Maintainers still have to review them.
So expect more advisories, more CVEs and more emergency releases, from more vendors, on shorter notice. The tools are also in limited hands today. Rob Lee at Dragos warns that the skill floor for the research that comes before an attack is dropping, and tools like this rarely stay limited for long.
What to change now
Shorten the browser cycle. Firefox and Chrome already update quickly, but many companies hold browser versions for compatibility testing. If an AI model can find 81 serious bugs in one go, holding back a browser for weeks is a bigger risk than an app breaking. Let browsers auto-update and test against the beta channel instead.
Rank by exploitability as well as CVSS. A larger flood of CVEs makes severity scores less useful as a queue. Use the CISA Known Exploited Vulnerabilities catalog and exploit-prediction scores to decide what goes first, and put internet-facing systems at the top.
Count your Linux kernels. The kernel results are the bigger story for servers, appliances and embedded devices. Many of those run old kernels that never get updated, because nobody owns them. Build the inventory now. You can’t patch a kernel you don’t know you have.
Plan for the vendors who are slow. Some suppliers will be swamped. Ask your key vendors how they’re handling AI-found reports and what their patch times are. If a vendor can’t answer, budget for compensating controls: segmentation, tighter access, and monitoring around that product.
Test the AI tools yourself. Defensive versions of these models are coming to security teams too. Running them on your own code and configurations before someone else does is the cheapest audit you’ll get.
The number to watch is Mozilla’s fix rate on the 81. If they’re all patched within a couple of release cycles, a well-run vendor can absorb AI-scale discovery. If they drag, assume every vendor you rely on will drag more.