Epoch AI researchers demonstrate AI-driven Firefox exploits

Researchers from Epoch AI have demonstrated that the open-weight AI model GLM 5.3 can autonomously exploit the “patch gap” to target Firefox users. This gap refers to the period between when a security fix is made public in the code and when it is actually deployed to end users.

In controlled experiments, the model successfully developed working file-stealing attacks against six specific Firefox releases. These attacks could potentially allow an attacker to read local files from a user's computer and send them to a remote server without further user interaction. The researchers noted that the API costs for developing these exploits were approximately $1,000 per version.

While the study involved an unpublished vulnerability that has since been patched, the findings highlight a significant capability for autonomous exploit development. The researchers designed the setup to simulate a low-resource attacker, providing only basic instructions to the model, which then worked independently to identify vulnerabilities and build attacks.