Open AI model GLM 5.3 hacks Firefox using unreleased security patches

Researchers at the non-profit organization Epoch AI have constructed an environment where the GLM 5.3 model was given information about a specific browser version and access to public code changes appearing after its release. The artificial intelligence autonomously identified security fixes and built multi-stage exploits.

• GLM 5.3 autonomously found unreleased fixes and built exploits within 12 to 48 hours.
• The attack worked on all six Firefox releases between September 17 and October 26, 2026.
• The sandbox protection was bypassed to read local files upon visiting a simple web page.
• The security flaw was patched in Firefox 159.0 on October 27 by the Mozilla team.

How the AI vulnerability attack worked

During the experiment, the AI model was supplied with specific browser version data and public repository updates. It autonomously identified security fixes, selected vulnerabilities, and constructed complex exploit chains capable of breaking out of the isolated rendering process to read arbitrary local files.

Attempts to replicate a similar attack against Google Chrome failed despite costing around $100,000, though researchers warn this does not guarantee complete protection. The core issue is the patch version gap, which allows malicious actors to weaponize fixes before end-users receive updates.

The danger of public repositories and open weights

The situation is exacerbated by open weights, which allow users to remove built-in safety guardrails and run the model without limitations. This research demonstrates that advanced cyber capabilities are no longer a monopoly of closed corporate laboratories.

Software developers are now urged to shorten the time window between publishing a patch in public repositories and delivering it to end-users, as well as avoid closing multiple vulnerabilities simultaneously to prevent exploit chaining.

What this means for the software industry

This milestone highlights a growing threat vector where AI models can weaponize developer transparency against themselves. For the broader tech industry, it emphasizes the critical need for rapid patch deployment and advanced defensive architectures that anticipate AI-driven vulnerability research.