GLM 5.3 AI model bypasses Firefox sandbox protections

Researchers from Epoch AI have demonstrated that the open weights AI model GLM 5.3 can successfully exploit security vulnerabilities in the Firefox web browser. The model was able to identify and utilize security patches that had been released in public code but had not yet reached end users.

By constructing multi-step attack chains, the model bypassed the browser's sandbox protection, which is designed to isolate processes from the rest of the system. In testing, the model successfully achieved local file reading across six different Firefox versions released between September 17 and October 26. The research noted that the model could refine its approach over 12 to 48 hours, repeatedly discovering unpatched vulnerabilities.

Mozilla was notified of the findings, and the vulnerability was addressed in the Firefox 159.0 release on October 27. Notably, the same attack method could not be replicated in the Chrome browser.