Google says hackers' AI now turns software fixes into attacks with no human involved
Google's security researchers say attackers have crossed a line defenders have long feared. In a report released Monday, Google's Threat Intelligence Group said it had seen hackers use AI to find and exploit software flaws with no human in the loop. Google says it is the first time its investigators have seen this.
For companies, the risk is not a new kind of break-in. Most attacks still start with an old weakness: a known flaw the victim has not yet fixed. What changes is the clock. If AI can turn a software maker's published fix into a working attack within hours, the weeks many firms take to install updates become a real exposure.
What Google found
The tool, which its operators named "exploit_agent," sat on an attacker's server that had been left open to the internet. It connects to several AI models, some commercial and some run on the attacker's own machines.
It compares popular web software before and after a security update. From the difference, it works out which flaw the update closed, writes an attack and launches it against sites that have not updated. Google says it does this automatically, step by step.
The same server held stolen victim data. It also held scripts for scouting targets, guessing passwords and planting hidden access points on hacked websites. Data on the server suggested the targets included French media, entertainment and education sites. Google did not name the group behind the tool.
Some caution is warranted. Google did not publish a success rate, and Hackread noted the report does not confirm that every attack the tool produced actually worked. The "first time" label describes what Google's own investigators observed, not proof that no one did it earlier.
The window to install fixes is shrinking
The report puts numbers on the pressure. So-called zero-days, flaws attacked before any fix exists, averaged eight a month in 2025, according to Security Affairs and UC Today. That rose to 11 a month this year and hit 22 in August. Disclosed flaws and attacks on them nearly doubled between January and August.
Security Affairs, summarizing the report, described a late-August case involving PaperCut, widely used office print-management software. Researchers, including the security firm Huntress, built working attacks on two of its flaws within 24 hours, the outlet reported. The threat-monitoring firm GreyNoise detected an attack campaign about four days after disclosure. In one break-in, according to Security Affairs, attackers went from a first foothold to top administrator rights on a company network within minutes.
AI also floods defenders with noise. After PHP, the programming language behind much of the web, fixed two flaws on October 21, hundreds of claimed attack demonstrations appeared. Most were AI-generated and many did not work. Security teams still had to check them, which BleepingComputer described as a significant extra burden.
Google expects this to get worse quickly. Within about three months, it expects simple attacks that need little or no human help, BleepingComputer reported. Within about six months, it expects "sophisticated, fully autonomous" attacks. The Decoder also reported Google's view that more powerful attacks are only months away.
What executives should take from it
Defenders still have one advantage. Google says these tools currently "trade stealth for velocity." Fast, automated attacks leave traces in network traffic and in the code they run, which makes them easier to spot. Google warns that this edge may fade as the tools improve.
Readers should also weigh who is sounding the alarm. Google sells AI-based security products, and UC Today flagged the company's own defense pitch as a vendor claim buyers should test. Some of the evidence, however, such as the PaperCut case, was also tracked by outside firms including Huntress and GreyNoise.
The practical steps are not exotic. Companies can measure how many days pass between a vendor's fix and its installation on systems that face the internet. They can ask software vendors and hosting providers how fast they patch, and who is responsible for which parts.
When a fix must wait, Hackread's summary of the guidance is to limit access to the affected application. Firms should also look for signs of an earlier break-in, because installing a patch does not remove an intruder who is already inside.
Florian Roth, head of research at Nextron Systems, told UC Today that the names of these AI tools matter less than their methods. "The names are disposable," he said.
For boards and insurers, the question is shifting. It is no longer only whether a company installs its security updates. It is how many days that takes, and whether that is fast enough against attackers who no longer need to sleep.