AI Turned Firefox Fixes Into Attacks in a Lab Test

In a lab test, GLM 5.3, an AI model anyone can download, turned Firefox's own public security fixes into working attacks on six Firefox releases. Each new release cost about $1,000 in AI fees, after more than $100,000 of research. That's Epoch AI's report of Oct 29. Nobody's browser was attacked: it ran on Epoch's own machine, cut off from the internet. Epoch says Firefox 159, out Oct 27, patches the hidden second bug the attacks needed, and it hasn't confirmed whether similar attacks still work there.