Google's security investigators say they have, for the first time in their tracking, found hackers using AI to find weaknesses in website software and use them in attacks with no person involved. In a report published on Monday, Nov. 16, the Google Threat Intelligence Group says it found the system, named exploit_agent, on a server the attackers had left exposed on the internet.
The system works from the security fixes that software makers release. In its previous report, Google had seen hackers automate only parts of this work. The system watches popular web apps and compares their code before and after a fix is announced, to spot the likely vulnerability the fix closed. A vulnerability is a weakness in software or a system that an attacker could use. A fix can show anyone who compares the code where the flaw was, and websites that have not installed it yet are still open. Then the system writes an attack that uses the weakness and launches it without a person stepping in, Google says.
It draws on several AI models, including some of the most advanced from AI companies and others the attackers run on their own computers. Google says systems like this can sharply cut the time between a fix being announced and attacks that use the flaw it fixed. That leaves website owners less time to install fixes. Google also says AI has set off a wave of attack code that is claimed to work but doesn't, which defenders must check.
Google says the same server held data taken from victims and scripts for later steps of a break-in, such as scouting a network and guessing passwords over and over. Others could plant webshells, small hidden programs that let an attacker control a website from afar. Records on the server suggest the operator was specifically interested in French websites in media, entertainment and education, Google says. Notes in the system's setup file are written in Russian, and they call the system a pilot.
Other tools automate parts of an attack rather than the whole of it. The report says hackers are using security-testing tools whose code anyone can download, which run teams of AI agents to automate steps of real break-ins. An AI agent is AI that can carry out tasks on its own, such as using apps and websites, instead of only answering questions. These ones look for simple weaknesses, write attack code as needed and adjust as they go, Google says. Besides ARTEX, a widely publicized tool, it names CAIRN, CyberStrikeAI and PentAGI. Google says it saw CAIRN and PentAGI used in real attacks, including CAIRN in break-ins whose targets included the U.S. aerospace industry, and saw hackers trying out CyberStrikeAI. Google says these tools work fast but leave traces that make them easier to spot and stop. It says they can still do what the attackers set out to do.
A group Google suspects is North Korean first reached cryptocurrency companies in North America and Europe through fake recruiter profiles with AI-made pictures, Google says. After breaking in, the attackers edited settings files for Claude Code, an AI coding assistant, in the companies' own code collections. When workers started the tool, it downloaded a backdoor, hidden software that let the attackers back in. The attackers changed the companies' own settings after they were already inside, and Google does not describe a flaw in the tool itself.
Criminals are also stealing companies' cloud logins to run AI at the victims' expense, Google says. Over 2026, the fraudulent charges in some cases passed $100,000 in a single day, and in the worst cases the total reached $1 million before the break-in was stopped.
The findings come from Google's own investigations. Google also offers businesses AI tools to defend against such attacks. Google does not say who ran exploit_agent, which AI models it used or how many websites the attackers broke into.