AI Model GLM 5.3 Autonomously Built Working Firefox Exploits From Public Fixes
According to researchers at Epoch AI, the freely available AI model GLM 5.3 turned publicly known security vulnerabilities into working exploits before updates reached Firefox users.
Researchers at Epoch AI report that the widely available AI model GLM 5.3 turned publicly known security vulnerabilities in the Firefox browser into working attack chains before the corresponding fixes reached users. According to the researchers, the model fully compromised six Firefox versions and did so without human intervention.
According to the report, each of these attack chains cost around 1,000 US dollars in API fees. None of the providers tested triggered their protective mechanisms. The work was carried out autonomously, and the complete exploits were produced within less than 48 hours.
The finding raises a fundamental question: security updates depend on attackers not being faster than the distribution of patches. If a freely available model can derive working exploits from published fixes in a short time, that window narrows considerably. The source describes the process as evidence that open models can accelerate security research and attacks in a single step.
The results come from an investigation by researchers at Epoch AI, which was picked up in its original form by the trade publication The Decoder. Further details on the test setup, independent verification, or the specific Firefox versions affected are not available. The available information also does not mention any response from the Mozilla Foundation or individual model providers.
For users in Germany, the finding above all underscores one thing: installing security updates remains crucial. The study shows that the gap between the disclosure of a vulnerability and its exploitation is shrinking once capable language models are used for this task. Whether and how providers adjust their protective mechanisms is likely to become an open question in the coming months.
7
