Epoch AI test: Zhipu GLM 5.3 turned unreleased Firefox security fixes into working attacks across 6 consecutive versions
In Epoch AI's tests, Zhipu's open-weight model GLM 5.3, across 6 consecutive Firefox versions, identified publicly available security fixes that had not yet reached the stable release as real patches and independently built complete attack chains, with no refusals throughout. The resulting web pages could send visitors' local files to the attacker without any additional clicks, and could reach saved passwords and cookies. The research team first invested more than $100,000 to build the security testing system; after that, each attack cost about $1,000, with the fastest taking 12 hours; each attack also used a second unknown Firefox vulnerability that the AI discovered on its own, which Mozilla fixed on October 27. A similar attempt against Google Chrome did not succeed even after spending about $100,000.
Why read this
SelectedAI score 85
Epoch AI's tests show that an open-weight model can independently turn unreleased Firefox security patches into working attack chains, along with the cost and time required per attack.
Topics
Body
In Epoch AI tests, Z .ai’s open-weight GLM 5.3 turned unshipped Firefox security fixes into working attacks (i.e. built working exploits) on 6 consecutive releases.
Those fixes sit in public code before reaching the stable browser, disguised as routine changes to slow attackers.
Without a single refusal from any provider, GLM 5.3 still picked out the real patches and built a full exploit chain alone.
The result was a webpage that sent a visitor’s local file to the attacker with no further clicks, a route to stored passwords and cookies.
The researchers first spent over $100K building a safe test system, but after that, each extra Firefox hack cost only about $1,000, and the fastest took 12 hours.
Every hack also used a second Firefox bug that nobody knew about, which the AI found by itself and Mozilla fixed on October 27.
The researchers tried the same thing on Google Chrome and failed, even after spending about $100K.