Open-weight model GLM 5.3 exploits Firefox patch-gap, costing ~$1,000 per attack

Epoch AI’s report demonstrates that the open-weight model GLM 5.3, released by Z.ai in August 2026, can exploit the “patch gap” in Mozilla Firefox link. The researchers built a webpage that, when visited, reads a file from the visitor’s computer and exfiltrates it without any further user action. Any Firefox version released between 17 September and 26 October 2026 was vulnerable, despite existing counter-measures that obfuscate security fixes as ordinary improvements.

The experiment required only publicly available information at the time of each version’s release and cost about $1,000 per successful attack. In addition to the patch-gap exploits, GLM 5.3 discovered an unpublished vulnerability, which Mozilla patched in Firefox 159.0 on 27 October 2026. The model faced no refusals from Z.ai when tasked with the exploit, and similar attacks on Google Chrome did not succeed at the time of writing.

Open-source projects are especially at risk because their codebases are publicly visible, but the authors warn that any software where upcoming security changes are observable could be vulnerable as open-weight cyber capabilities improve.

Our read: The findings prove that widely available models can lower the expertise and cost barriers for exploiting timing windows in software updates, forcing enterprises to accelerate patch deployment and consider AI-driven threat modeling as a core defensive practice.