Google warns AI-assisted hacking found

Cybersecurity people have been warning about this for a while. The scary version was never “AI writes a spam email faster.” It was “AI helps find and weaponize a bug nobody knows about yet.” Google now says that line has been crossed. On May 11, its Threat Intelligence Group said it found the first case it believes a threat actor used AI to develop a zero-day exploit, and that the exploit was headed for a mass attack before Google interrupted it. ### What is the new thing here? A zero-day is a vulnerability nobody has patched yet, because defenders do not know it exists. That already makes it dangerous. The new part is not just that criminals used AI somewhere in the workflow. Google says the AI likely helped develop the actual exploit — the working attack code for a previously unknown flaw. That is the milestone. (cloud.google.com) ### What did the exploit do? Google says the bug let attackers bypass two-factor authentication in a popular open-source, web-based system administration tool. Google did not publicly name the product, but the capability matters more than the brand here — 2FA is supposed to be the extra lock on the door. If attackers can step around it, a lot of normal defensive assumptions break fast. (cloud.google.com) ### Was this used in the wild? Not at full scale, apparently. Google’s line is that the criminal actor planned to use the exploit in a mass exploitation event, but Google’s “proactive counter discovery” may have stopped that before rollout. That distinction matters. This was not a postmortem after a global incident. It was a disruption before the campaign fully landed. Still, the existence of the exploit is the real warning. (politico.com) ### Why does AI change the threat? Because exploit development is usually slow, specialized work. You need to inspect code, test edge cases, write proof-of-concept logic, and keep iterating when the first version fails. Generative models can now help with that loop. Basically, they act like a tireless junior reverse engineer that never gets bored. Not perfect, but fast. Google says attackers are using AI to increase the speed, scale, and sophistication of operations, not just to save a few hours. (cloud.google.com) ### Is this just one weird case? Probably not. Google frames this as part of a broader shift from experimental AI use to industrial-scale use across adversary workflows. Its report also points to China-linked and North Korea-linked actors exploring AI for vulnerability discovery, Russia-linked actors using AI-augmented malware techniques, and autonomous malware that can generate commands based on what it sees on a victim machine. That does not mean every hacker is suddenly superhuman. (politico.com) But it does mean the toolchain is getting better. ### Does this mean frontier models caused it? Not necessarily. One detail that got attention is what Google did not say. Politico reported Google concluded Anthropic’s Claude Mythos was most likely not the model used here. So this is not a clean story of one famous frontier model directly powering one criminal exploit. The broader point is worse, in a way — attackers may not need the very best public model to get useful exploit help. (cloud.google.com) ### What does defense look like now? More aggressive vulnerability hunting, faster patching, tighter access controls, and better tracing of how tools connect into internal systems. Google is also pushing the other side of the AI equation — using AI systems like Big Sleep and CodeMender to find and fix bugs faster. That is the race now: machine help for attackers versus machine help for defenders. (politico.com) ### Bottom line? The headline is not that AI has replaced hackers. It has not. The headline is that AI has apparently helped clear one of the hardest steps in offensive cyber work — turning an unknown flaw into a usable exploit. Once that becomes routine, the economics of hacking change. (cloud.google.com) (blog.google)