AI Model Exploits Firefox Patch Gap Before Fixes
Researchers say the patch gap, the window before fixes reach users, can be exploited by AI models, with Firefox 159.0 now closing the flaw.
Updated on Nov. 2, 2026 in Cybersecurity
Live Poll
Do you feel safe using your browser given how fast AI can exploit updates?
Researchers used the GLM 5.3 model, released by Z.ai in August, to build working Firefox exploits from security patches that had not yet been shipped. The attacks succeeded against all six Firefox versions released between September 17 and October 26, 2026, and Epoch AI published its report on October 29.
Why it matters
The work shows that a patch gap, the window between a fix being developed and it reaching users, can be exploited by an AI model that reads only publicly available information about a release. It also suggests that defending against such attacks may depend on how quickly fixes reach users rather than on keeping patch details secret.
The evidence comes from controlled experiments in which the model was given one Firefox version and only the public information available when that version shipped. Each successful run chained two flaws and cost about $1,000, while the same approach against Chrome produced no working exploit despite roughly $100,000 in API spending; whether Chrome resists this approach for lasting reasons or only because of the model's limits is not yet established.
The players
Z.ai
Chinese AI developer that released the GLM 5.3 model in August, which was used to build the exploits.
Mozilla
Maker of the Firefox browser, which has more than 200 million active users and fixed the sandbox flaw in Firefox 159.0.
Epoch AI
AI research organization that published the report on the exploits on October 29, 2026.
The details
The researchers gave GLM 5.3, a large language model released by Z.ai in August, a specific Firefox version and only the information that was public when that version shipped. The model diagnosed crashes, discarded failed approaches and chained several stages together until visiting a malicious web page was enough to compromise the browser. Each attempt was capped at 48 hours, and the fastest success took 12 hours. A second, unpublished vulnerability was used to escape the browser's sandbox, the isolated environment meant to contain malicious code.
Timeline
September 17 to October 26, 2026: Six Firefox versions are affected by the working exploits.
October 27, 2026: Mozilla fixes the sandbox flaw in Firefox 159.0.
October 29, 2026: Epoch AI releases its report on the exploits.
The Tech Race
The results sit alongside a broader contest over whether AI models can find and weaponize software flaws faster than vendors can ship fixes. The Chrome result, in which the same method failed at about $100,000 in spending, is the clearest comparison point, and the report says more capable open models such as Kimi K3.1 Pro could perform better against Chrome.
Firefox users are protected once the Firefox 159.0 fix reaches them, which the report says is expected to make similar attacks much harder. The report does not give a timeline for Chrome updates or for any further patch to the unpublished sandbox-escape flaw.
The takeaway
The central lesson is that the interval between a patch being written and users installing it is itself an attack surface, and AI-assisted exploitation shortens the time needed to use it. Watch whether Mozilla or Google publicly confirm patches for the remaining sandbox flaw and whether follow-up tests against Chrome with newer open models produce a working exploit.
Further reading
For background on how these threats are tracked and countered, see Cybersecurity coverage.
Source note: This article includes information reported by Teknoloji Haberleri - ShiftDelete.Net.
Live Poll
Do you feel safe using your browser given how fast AI can exploit updates?
