Google Exposes AI Hacking Framework That Turns Patches Into Working Exploits

AI is now writing the playbook for break-ins, and Google has laid out how far it has gone. Google Threat Intelligence Group, working with Mandiant, published the latest AI Threat Tracker, documenting multi-agent intrusion tools and an exposed system that automated patch analysis and exploitation attempts.
The system at the center of the report is called “exploit_agent”. Google says it took patches, worked out what they fixed, and turned that knowledge into working exploits. That flips the usual race on its head: the same patch meant to close a hole becomes the map to it.
What exploit_agent did
According to Google’s AI Threat Tracker, exploit_agent automated the slow, skilled part of hacking — reading a patch, finding the flaw behind it, and building an attempt to use it. The report links the system to a suspected Russian-speaking operator, and it sits alongside other multi-agent intrusion tools that split the work of an attack between several AI agents.
That division of labor is what makes the trend hard to stop. One agent can scout, another can probe, another can act — and the whole chain runs faster than a human team would.
The FBI’s warning on Artex
Google’s report is not the only alarm. The FBI separately warned that the open-source Artex AI tool helped hackers break into real networks in under three hours. The cases cited include a U.S. IT company and targets in South Korea’s financial sector, a sign that these tools are already being used against live systems, not just tested in labs.
Artex shows the other side of the same shift. Where exploit_agent was exposed by researchers, Artex is open source — out in the open, ready for anyone to pick up.
What defenders should take from this
Patching fast matters more than ever when the patch itself can be turned into an exploit in hours. Defenders need to assume that a fix will be read, reversed and reused, and to plan their rollout with that in mind.
Supply-chain attacks are already hitting developers hard, as the Cursor breach showed, and Microsoft has linked AI agents to backdoors planted by North Korea’s Lazarus Group. Mandiant, which worked with Google on the tracker, has long tracked these groups — and the message from this report is that the tooling has caught up with the intent.
