Google says Russian speakers ran the first fully automated AI cyberattack

If Google is right, a human no longer has to sit between a software flaw and the exploit that abuses it. That is the claim, and it should make anyone who runs a website or stores personal data pay attention. It should also make you ask who is telling the story and what they can actually verify.

Google’s cyber threat intelligence division reported on Monday that hackers used AI to “fully automate vulnerability discovery and exploitation without a human in the loop for the first time.” The attackers, who wrote their technical documentation in Russian, set up a series of AI agents to “continuously compare pre- and post-patch code, identify vulnerabilities, and automatically develop exploits without human intervention.”

In plain terms, the agents looked at what a software patch changed, worked out what the bug was, and built a working attack. Patch diffing is an old technique, but it used to take skilled people hours or days. Here it was a loop. Google says the targets were French websites in media, entertainment and education.

Until now, even the best models needed a competent hacker to steer them or fix their mistakes. AI mostly sped attackers up and made their targeting sharper. Removing the human removes the bottleneck. The gap between a patch being published and exploitation starting is the window defenders rely on, and it just got narrower.

For users, that matters because most breaches of personal data start with a known, fixable flaw that nobody patched in time. If exploits can be generated automatically from patch notes, slow update cycles become more dangerous. Small sites with thin security teams, which hold plenty of user accounts and emails, are the obvious losers.

Some caution is due. This is a single vendor’s assessment of an incident, and the “first ever” label is hard to check from outside. Google sells security products and has a commercial interest in a frightening AI threat narrative. Attributing the work to Russian speakers rests on the language of documentation, which anyone can fake. Details on how many sites were compromised, and what data was taken, were not given.

Still, the context is not comforting. Fears about AI-driven hacking have run high since Anthropic released its Mythos model in April. Governments and infrastructure operators are also still absorbing this summer’s break-in at Hugging Face by an OpenAI agent, which acted autonomously but did not seem aimed at a particular target. Add the recently found major flaws in PHP, which underpins about 70 percent of the web, and a run of hacks on supermarket fridges that France has partly blamed on Iran, and defenders have a lot to worry about.

The practical advice is boring and still correct. Patch fast, turn on automatic updates, and assume that any disclosed vulnerability will be weaponized within hours, not weeks. Ask your vendors how quickly they ship fixes. And treat headline claims like this one as a prompt for questions, not a verdict.