Skip to content
Brief Security & Trust ·

Google says attackers ran an AI tool that turns patches into exploits

Google's threat team says no person steered the tool, a first in attacks it has seen.

In brief
  • Google's threat team says attackers ran an AI tool on its own to turn patches into exploits. Two months ago it had seen no such attacks.
  • In a separate case, AI models produced working exploits within a day of a PaperCut disclosure. WebPulse suggests asking how fast your critical patches go live.

Google's threat intelligence team said on Monday that attackers have, for the first time, run an AI tool without a person steering it. SiliconANGLE reported the findings. The tool, called exploit_agent, reads a popular web app's code before and after each fix. It spots the change, drafts an exploit for the flaw, and fires it automatically. Google found it on an exposed staging server linked to several language models.

The server also held data taken from victims, plus brute-force scripts and tools for planting webshells, which are hidden remote-control pages. Google said victim data points to French media, entertainment and education sites. Two months ago, it had reported no such fully autonomous attacks. SiliconANGLE's account does not say who runs the tool. Separately, researchers and Huntress Labs got working exploits from AI models within a day of two PaperCut flaws being disclosed.

This part is WebPulse's reading, not Google's. A patch shows where a flaw lives, because the changed code points to it, and the PaperCut case shows that gap can be a single day. Ask your team how fast critical patches go live, which internet-facing apps are exposed, and what your vendors promise on patch speed.

A WebPulse Brief: a short report of an important event, written by the WebPulse Newsroom with AI assistance and checked against the reporting below. When there is more to explain, we follow up with a full story. How we use AI.

Reporting: SiliconANGLE.