Two-factor authentication is the security step most people recognize as the code sent to your phone after you type your password. Businesses use it because even if a criminal steals your password, they still cannot get in without that second code. It is one of the most widely recommended security practices in the world. This week, Google confirmed that a criminal group used AI to find a hidden flaw in a popular administration tool that would have let them skip past this protection entirely, without needing to trick any user.
This is the first time a major security group has confirmed an AI-built attack of this kind in the wild. The distinction matters. AI has been used before to help criminals write faster phishing emails or tidy up malicious code. What happened here is different: the AI appears to have found a vulnerability that no one knew existed, and then built the tool to exploit it. Google stopped the group before they could launch a mass attack.
How did Google know AI was involved? The code the criminals wrote had signs that a human would not leave behind. It included a detailed help menu, a highly structured format, and a security score for the flaw that was simply made up, a number that AI systems sometimes invent when they do not have a real one to hand. Human hackers writing an exploit tool do not add unnecessary documentation. AI does, because that is how it was trained to write code.
The flaw itself is important to understand. The software had a hidden contradiction built in by its developers: the two-factor check was turned on in one part of the code, but another part quietly told the system to skip it under certain conditions. A human scanning for vulnerabilities would likely miss this because it requires reading the intent of the code, not just looking for crashes or obvious errors. AI turns out to be surprisingly good at this kind of reading, and that changes the threat model for every business that runs software.
North Korean state hackers sent thousands of automated queries to AI tools to build a library of working exploits. Chinese groups tried to trick Google's own AI into acting as a senior security expert so they could extract vulnerability information. Russian-linked groups used AI to generate large blocks of dummy code that hid their real malware inside, making it much harder for standard security tools to detect.
The broader picture from Google's own data is that almost half of all serious software vulnerabilities tracked in 2025 were aimed at businesses, the highest share ever recorded. The average data breach now costs a company around $4.44 million globally, though US companies average more than double that. And small and mid-sized businesses are increasingly the preferred target, accounting for the majority of breaches in 2025, precisely because they tend to have weaker defenses.
The criminal group in this case planned a mass exploitation event, meaning they intended to hit many targets at once. That is the normal playbook for financially motivated cybercriminals. They find one flaw, build one tool, and run it across thousands of victims before anyone patches the hole. AI shortens the time it takes to get from "found a flaw" to "ready to attack."
Google's chief analyst said plainly that for every AI-built attack they can trace, there are probably many more they cannot. The attack that was stopped this week had telltale signs of AI involvement left in the code. A more careful criminal group would clean those up. The next one may leave no obvious trace at all.
The security practices that businesses have relied on for years, passwords plus a second code, regular software updates, basic employee training, are not wrong. But they were designed for a world where finding a new, hidden vulnerability took skilled human researchers weeks or months of work. That timeline is shrinking fast, and the defenses that businesses invest in need to account for that.