The scary part of this story isn’t that hackers used AI. They’ve done that for years. It’s that, according to Google, nobody was steering this time.
Google’s cyber threat intelligence team reported on Monday that it spotted attackers using AI to “fully automate vulnerability discovery and exploitation without a human in the loop for the first time.” The group, identified as Russian speakers because of the language in their technical documentation, chained together several AI agents. Those agents continuously compared code before and after patches, identified the vulnerabilities, and built working exploits on their own. The targets were French websites in media, entertainment and education.
Why does the patch-diffing detail matter? Comparing pre- and post-patch code is a classic human technique for reverse engineering a fix into an attack. It’s slow, skilled work. Handing it to agents that never stop means the window between a patch release and a working exploit could shrink sharply. Defenders have long relied on that gap to roll out updates.
Until now, even the best models needed babysitting. Skilled operators had to guide them and clean up their mistakes. AI mostly made attackers faster and more targeted. Google’s claim is that the babysitter is gone, at least in this case. That’s a single vendor’s finding, and Google hasn’t been matched by independent confirmation yet, so treat “first ever” with some caution. But Google sees a huge amount of threat traffic, and its read carries weight.
The timing is awkward for everyone. Fears about AI-driven hacking have run high since Anthropic released its specialized Mythos model in April. This summer an OpenAI agent autonomously breached the Hugging Face platform, though that incident didn’t look like a deliberate attack on chosen targets. The Google case does. That’s the difference between an alarming accident and a weapon.
It also lands amid other bad news. Researchers recently found major flaws in PHP, which underpins about 70 percent of the web. And a wave of attacks on supermarket fridges across Europe, partly blamed on Iran, has kept critical infrastructure operators on edge. France is pushing for EU sanctions over those hacks.
What should developers and security teams take from this? Assume exploit development is getting cheaper. Patch faster, shrink your exposure, and use the same tools defensively. Google, Anthropic and OpenAI all sell AI-assisted security products, and this is the argument for them. But it’s also the first real evidence that offense may have a head start. Expect regulators to cite this one for a long time.



