Open-weight AI model exploits Firefox flaws in 48 hours with no human input

Epoch AI researchers report that a freely available Chinese AI model, GLM 5.3, exploited Firefox security patches before they reached users. In tests, the model autonomously compromised six Firefox versions within 48 hours at a cost of about $1,000 per successful attack in API fees. No API provider detected or blocked these attempts during the experiments.

The model developed autonomous exploit chains, identifying vulnerabilities in published patches and bypassing browser protections. It exploited an undisclosed security flaw to read local files on user machines. The fastest attacks took 12 hours, with a 48-hour cap for the model. Researchers note that Firefox disguised security fixes as minor improvements, without success.

This attack exploits the "patch gap", the delay between patch publication and user deployment. Researchers warn that the entire software industry is vulnerable, including mobile systems and router firmware. They recommend limiting access to patches until deployment and accelerating updates, while avoiding bundling multiple fixes to reduce exploit chaining risks.