Open-weight model GLM 5.3 built full Firefox exploit chains with zero human input in under 48 hours

Why it matters

Epoch AI researchers demonstrated that GLM 5.3, an openly available AI model, can autonomously exploit Firefox by abusing the "patch gap" — the window between a security fix becoming visible in code and reaching users. Working offline with only basic instructions, the model identified disguised security fixes, chained vulnerabilities, and built complete exploit chains without human input, ultimately bypassing the browser sandbox to read local files from anyone visiting a crafted webpage. The researchers note that no API provider triggered safety guardrails during the work, and they describe their setup as approximating what a low-skill attacker could achieve. Mozilla patched the underlying flaw after disclosure, though the researchers have not verified whether attacks still work. The team warns the patch gap affects the entire software industry, especially open-source projects and infrequently updated software like router firmware and mobile operating systems, and argues that faster updates help but cannot close the structural problem. Anthropic had previously reported comparable exploit capability from Mythos.