AI news Oct 30, 2026

Today

In an Epoch AI experiment, GLM 5.3 worked from public but unshipped Firefox fixes, plus a separate bug GLM 5.3 found, to build attacks on six releases at about US$1,000 each

Why it matters: Our inference: Firefox fixes are public before most users have them, and Epoch's model turned that gap into a working attack for about US$1,000 a release. That price also needed a separate bug the model found, since fixed, and Epoch hasn't confirmed the gap alone is as cheap.

Epoch AI, an independent research group that tracks AI compute, costs and capabilities, published an experiment on October 29 in which an AI model attacked Firefox.

The experiment measures what it costs to turn the patch gap into a working attack. Firefox fixes appear first in the publicly downloadable beta and nightly builds and reach ordinary users some time later. During that window anyone can see what was fixed, while most people's browsers haven't updated. Epoch had GLM 5.3, the open-weight model from China's Zhipu, start from those public fixes and attack six Firefox releases shipped between September 17 and October 26. The model produced a working exploit for all six. Epoch says that on those six releases, a user only has to open a web page, with no further clicks, for an attacker to steal the passwords and login credentials stored in the browser or read files on the computer. Each release cost about US$1,000 to break. The model got at most 48 hours per attempt, and the fastest took 12 hours (Epoch AI on X, 2026-10-29; Epoch's full report, 2026-10-29).

What is new is the raw material and the price. Epoch writes that exploiting the patch gap used to take a lot of skilled labor, or access to a frontier model with advanced cyber capabilities. Epoch gives no cost for those earlier approaches, so there is no way to say how much cheaper US$1,000 is. Anthropic measured something different on September 29: on Anthropic's internal set of 100 binary-exploitation problems, GLM-5.3 built a complete exploit on 4% of them, against 6% for Claude Mythos Preview, an Anthropic model that hasn't been released (Anthropic, 2026-09-29). Epoch's experiment instead measures the cost of breaking a Firefox release that is already in users' hands.

Epoch's full report lists three limits on the result. On recent Firefox releases the attack did not achieve arbitrary code execution on the victim's machine; success meant reading files and sending them out. To get from the browser to the files, GLM 5.3 found and used a previously undisclosed vulnerability during the experiment. Mozilla fixed that vulnerability in Firefox 159.0 on October 27, and Epoch says it hasn't confirmed whether the attack still works without it. An attempt on Chrome produced no attack of the same grade. The US$1,000 is the marginal cost of one more release; the whole study cost more than US$100,000.

Not proven yet: Epoch hasn't confirmed the result holds on Firefox 159.0 and later How we checked →

Next to watch: whether Epoch publishes a retest on Firefox 159.0 and later, and whether another team repeats the patch-gap attack on other software. Firefox 159.0 has already fixed the bug the model used. A retest that fails within the same 48 hours would mean the US$1,000 price depends on the model finding an extra bug.