On October 29, 2026, Epoch AI reported that GLM-5.3, the open-weight model Z.ai released in August, could exploit the “patch gap” - the window between a security fix being published and reaching users. Working autonomously from a basic instruction, the model built a webpage that, when visited in Firefox, read a local file and sent it to the attacker with no further user interaction.
All six Firefox releases tested (156.0 through 158.0.1, released September 17 to October 26) were compromised. In every case the model also used an unpublished vulnerability alongside the patch-gap flaws; Mozilla has since fixed it in Firefox 159.0. The marginal cost was about $1,000 per version at API pricing, attempts were capped at 48 hours and the fastest success took about 12 hours. Total Firefox research spend exceeded $100,000.
The model ran on Ubuntu 26.04 with default security settings and official Firefox builds, with no internet access and only information available at each release date. The task was framed as a security exercise and the researchers saw no refusals or safeguard triggers. The same harness spent about $100,000 on Chrome without an equivalent exploit.
Why it matters: it is a concrete, priced demonstration that a freely downloadable model can weaponize published fixes faster than many users install updates, which shortens the time defenders have between patch and deployment.
What it does not show: the exploits achieved only file reads, not arbitrary code execution; they depended on a since-patched unpublished bug and were not confirmed on 159.0; only six releases were tested; the Chrome failure does not rule out better prompting or stronger models; and the authors draw no conclusion about the overall offense-defense balance of open-weight models.