
For a couple of years now, the warning has been the same: at some point an AI will carry out a cyberattack entirely on its own, with no human steering it. Google says that point has arrived. On 17 November its threat intelligence group reported that a suspected Russian-speaking operator used an AI agent to find software flaws and write working exploits for them without a human in the loop, which it described as a first of its kind. The line everyone was watching for has been crossed.
What actually happened
The clever and chilling part is how ordinary the raw material was. When a software company issues a security patch, it is effectively publishing a map of the hole it just fixed. Skilled attackers reverse-engineer those patches to work out the vulnerability and then build an exploit for anyone who has not updated yet. That work normally takes a capable human some time. According to Google, an AI agent did the whole chain, understanding the patch, identifying the flaw, and producing a working exploit, by itself. Separately, the FBI has issued an advisory noting that since at least August 2026 a range of malicious actors have used an offensive AI tool called ARTEX to run campaigns against organisations worldwide, which readers of this site will recognise from earlier reporting.
Why this is the moment that matters
Up to now, the limiting factor in offensive hacking was people. There are only so many skilled operators, and each one can only do so much. That human bottleneck is what kept the volume of serious attacks within bounds. Remove it, and the maths changes completely: one operator can run thousands of simultaneous attacks, around the clock, for the cost of compute. This is the exact dynamic that turns cheap, deniable sabotage into a weapon of hybrid warfare, and it is why a single reported case is worth paying attention to even before it becomes common.
The honest caveats
Hold two things in mind before you panic. This is one reported incident, the attribution is described as suspected rather than confirmed, and the company doing the reporting, Google, is also selling the defence. Google used the same announcement to pitch its new AI Threat Defence, which stitches together its Gemini models, the security firm Wiz, an automated patching system and its Mandiant unit to find and fix flaws before attackers reach them. A vendor warning that the sky is falling while holding an umbrella for sale deserves a raised eyebrow. But the direction of travel is not in doubt, and the people who do this for a living have been expecting exactly this.
The uncomfortable conclusion
If attacks can now run themselves, defence has to as well. A Google executive put it plainly: rely on manual effort and you fall behind, so the answer is agentic defence, AI guarding the gates against AI storming them. That is the future this incident points to, and it is a strange one. The cat-and-mouse game that has defined cybersecurity for decades is losing its humans on both sides, leaving two sets of machines fighting at a speed no person can follow, with the rest of us downstream of whoever’s model is better that day.
What this means
Treat the specific claim with appropriate scepticism and the trend with none. The fully autonomous cyberattack has gone from a thing security researchers warned about to a thing one of the biggest companies in the world says it has seen. The only defence anyone is seriously proposing is more AI. We spent two years asking when the machines would start hacking on their own. The answer, apparently, is now.
Edgar Friendly, calling the AI bullsh*t. More tools, more swearing: toptoolstack.com.
Did you know? A security patch is a double-edged thing. It fixes a flaw for everyone who installs it immediately, and it quietly advertises that same flaw to everyone who has not. AI now makes reading that advertisement and acting on it a job that needs no human at all.
Sources: NewsNation; The Hill; Ground News.