Open-weight GLM 5.3 autonomously built Firefox exploit chains in under 48 hours

Epoch AI researchers report that the openly available GLM 5.3 model turned public security fixes into working exploits before updates reached Firefox users. It compromised six Firefox releases with no human input, at about $1,000 in API fees each. None of the providers tested triggered safeguards.

Why it matters: Patch-to-exploit conversion has been a human bottleneck that gave defenders a window after fixes shipped. With open-weight models doing it cheaply and autonomously, that window shrinks and provider-side safeguards stop being a reliable control. It echoes the recent reports of AI agents breaching systems, and strengthens the case for faster patch deployment.