An AI Hacking Tool Hit 200 Organizations. It Skipped the Protected Ones.
Researchers say a state-linked group pointed an AI pentesting framework at 14 countries, and the operators told it to avoid machines with endpoint protection.
I'll get to which computers the hackers avoided. Stay with me, because that one detail is the most useful thing you'll hear tonight.
What happened
The security company CrowdStrike tracks a China-linked group it calls Vapor Panda. Starting October 17, that group went after more than 200 organizations in 14 countries. Government ministries. Telecom companies. Drug makers. Chip companies. Defense. Nuclear. Finance.
And they didn't do it by hand. They used an open-source tool called ARTEX AI. Picture a team of software agents. One scouts the network. One finds a weak spot and pushes on it. One plants a back door. One grabs passwords. And if a path is blocked, they try another one. No coffee breaks.
Researchers confirmed at least nine organizations were actually compromised, including telecom providers in the U.S. and Belgium, government organizations in Greece, the Middle East and Latin America, and pharmaceutical firms in the U.S. In one intrusion, researchers estimate the AI agents chewed through more than 1.4 billion tokens of input. Tokens are the chunks of text these models read, roughly pieces of words.
Humans still held the steering wheel
People were still in charge. Operators gave the agents instructions in Chinese, ran several break-ins at the same time, and stepped in at the big decision points.
The detail I promised you
And here it is. The operators told the AI to skip Windows machines protected by CrowdStrike's Falcon security software. Go around them. Hit the unprotected ones instead.
Think about what that means. Even an AI-powered attacker, working at machine speed, chose to walk around the computers that had real protection on them. Defense works. It raises the cost. Attackers go where it's cheap.
How fast this is now
One more data point. A Dutch security research group, DIVD, was hit by the same tool in September. It went from first foothold to full control of the machine in twelve minutes. Twelve. DIVD counted more than 450,000 requests and 2,500 scripts tied to the attack, and called it, quote, "super loud and messy."
So the attacks are fast, and they're noisy. That's actually good news for anyone who's watching their logs.
What you do with this
Same advice as yesterday, because it keeps being true. Keep security software on your machines and keep it updated. Turn on two-factor. Don't reuse passwords. You don't need to be impossible to hack. You need to be more annoying to hack than the next guy.