Security companies have warned for a couple of years that hackers would eventually hand the whole job to AI. On Monday, Google said it had seen it happen.
The finding comes from the latest AI threat report by Google's threat intelligence team and Mandiant, the company's incident response unit. It describes a tool the hackers named exploit_agent. Google says this is the first time it has seen attackers use AI to find software flaws and break in through them "without a human in the loop."
The story spread fast. A post from unusual_whales passed 3.8 million views overnight, and posts from TMLwire and Osint613 added close to a million more. One line traveled with it: "the era of autonomous attacks has arrived." We couldn't find those words anywhere in Google's report or its public statements. What Google actually found is more limited than the posts suggest, but it's still worth worrying about.
What Google found #
Google didn't catch the tool in the middle of an attack. Its researchers found it on a server the hackers used for testing and had left open to the internet. Next to it were data stolen from victims and the tools hackers use once they're inside a company's network.
When a software company finds a security hole, it sends out a fix, usually called a patch. That patch also shows anyone paying attention where the hole was. exploit_agent is built to take advantage of that. A team of AI programs watches popular software for new patches. When one comes out, the programs compare the old code with the new code, work out what the fix repaired, write an attack aimed at that hole and launch it at computers that haven't installed the update. Nobody has to approve any of it.
The server also held more routine hacking tools, for mapping a network, guessing passwords and planting hidden back doors. One of them deleted back doors that rival hacking groups had left on the same machines.
What Google didn't say #
Google hasn't said who built the tool. CyberScoop described the hackers as Russian-speaking, and the evidence for that is a settings file with comments written in Russian. The file calls the project a pilot. That's a reasonable guess, but it isn't proof.
VentureBeat pointed out some bigger gaps. Google doesn't name a single flaw the tool used. It doesn't say how many victims there were or how quickly the tool got from a new patch to an attack. It also doesn't show that this tool caused the break-ins whose stolen data was found next to it. The victim data points to French news, entertainment and education websites, and one code sample shows the tool testing a WordPress add-on called Spectra.
It helps to remember who's raising the alarm. Google and Mandiant both sell security services. That doesn't make the report wrong, but calling this the "first" is Google's own judgment.
Why it matters anyway #
What stands out is how fast this has moved. In September, Google told SiliconANGLE it had not yet seen hackers use fully automated attack systems on real targets. Now it says it has. In May, an earlier edition of the report described what Google believed was the first "zero-day" attack built with AI help. (A zero-day is a flaw that attackers use before the software maker knows about it or has a fix.) In about six months, AI went from helping hackers to doing the job for them.
Other numbers in the report point the same way. Google counted 22 zero-day flaws being used in attacks in August. Earlier this year the figure was about 11 a month, and in 2025 it averaged about eight. When two flaws in PHP, a programming language that runs a large share of websites, were patched on Oct. 21, hundreds of attack programs claiming to use them showed up within days. Most didn't work, and AI probably wrote them. The quality isn't there yet, but the volume already is.
Google expects this to keep growing. These tools are cheap to run and easy to update, and a small crew can point them at a huge number of targets.
The FBI warning #
The FBI warning that went viral Tuesday morning is related, but it isn't new. The alert, FLASH-20261113-001, came out last Friday, and the FBI said Tuesday it hasn't issued anything since. It's still worth reading, because it shows what AI-driven attacks look like in real life.
The alert is about ARTEX, a free tool that security teams use to test their own systems for weak spots. A developer who goes by Autumn-27 posted it on GitHub in July. By August, criminals were using it against real companies. The developer took it down in early October, but that didn't stop it. Copies and a paid version called ARTEX PRO are still out there, and in early November researchers found more than 1,600 ARTEX control panels open on the internet, BleepingComputer reported.
The FBI links ARTEX to attacks on South Korean financial firms that exposed personal data on about 67,000 people. It also links the tool to a break-in at ConnectWise, which sells IT software to other businesses. That break-in opened a path into more than 2,250 of its customers' systems. In some cases, attackers went from first contact to high-level access inside a network in under three hours.
One detail in the alert deserves more attention. ARTEX can run on well-known AI models like Anthropic's Claude and OpenAI's GPT, but most of the attackers the FBI tracked used cheaper Chinese models, mainly DeepSeek's. The FBI says those models cost less, work well enough and rarely refuse a request that's dressed up as security testing. Even if the big American AI companies block misuse, attackers have other options.
Neither ARTEX nor exploit_agent has been linked to last week's Microsoft outage or Friday's nationwide flight ground stop, which officials traced to a router problem. With so much breaking at once, it's easy to assume an AI attack is to blame. So far, there's no evidence of that.
Why update timing matters now #
Most companies don't install security updates the day they come out. An update gets tested, scheduled and rolled out during a planned maintenance window, and that can take days or weeks. Until it's installed, the company is exposed. That used to be an acceptable risk, because turning a patch into a working attack took a skilled hacker several days. Merritt Baer, a former deputy chief information security officer at Amazon Web Services, described the change to VentureBeat: "Now an agent does it while its operator sleeps."
exploit_agent is designed for exactly that window. It can start working the moment a fix is published, often before a company has even put the update on its schedule. In another case cited in Google's report, as CSO Online reported, the security firm GreyNoise found that attackers likely used AI to build attacks on PaperCut, a widely used printing program. They hit hundreds of organizations four days after the flaw was announced. Plenty of companies take longer than that just to approve a patch.
Companies are using AI to defend themselves, too. Bank of America CEO Brian Moynihan recently told analysts, according to Business Insider, that the bank had hired 1,000 outside consultants to fix software flaws. He called the effort "Mythos remediation," a reference to Anthropic's Mythos AI models. Both sides now have the same kind of tools. For now, the attackers are moving faster.
Old advice, shorter deadline #
The advice coming out of this week isn't new. The FBI's alert starts where security people always start: know every system you have that can be reached from the internet, including the forgotten ones, like an old partner portal or an app nobody has touched in years. Those are the first places an automated attacker will look.
What has changed is how fast companies have to act. Amit Kumar Jena of the tech consultancy Kanerika told CSO Online that internet-facing systems should get their own fast track for updates, with temporary protections in place while the real fix is tested. Keith Prabhu of Confidis said companies should also decide ahead of time which emergency moves can run on their own, such as cutting a laptop off the network, and keep human sign-off for the big decisions.
That last part may be the hardest sell. At a lot of companies, a patch still needs a meeting before it goes in. The tools Google and the FBI described don't wait for meetings.
Google's researchers did leave one opening. These AI tools are noisy, because the people running them care more about speed than about hiding. A company that's actually watching its systems has a decent chance of catching them. One that isn't will probably hear about it from someone else.