AI Agents Breached 100+ Companies for $50 a Hack

NorForge researcher Snorre Fagerland found more than 20 exposed control panels for a system called Cairn. Cairn runs AI agents without a human in the loop, and the agents hunt for vulnerabilities in websites and break in on their own. One operator pointed the agents at over 1,200 sites and breached more than a hundred of them, including Deere, SharkNinja, and Cummins. Each successful breach cost about $50 in tokens. The agents' work was signed under the name claudecode_deepseek-v4.1-flash. They ran unsupervised against a single target for almost a full day, 23 hours on average.

For this kind of mechanical work — running the same vulnerability-scanning script around the clock — a cheap model like DeepSeek is good enough. The task is narrow and repetitive. It does not call for real engineering judgment, which is exactly where these cheaper models tend to fall short. My own Claude Code agents work on a similar principle: I hand the hard problem to Opus or Fable, and the tedious trial-and-error to a cheaper model. The real lesson here is not about an AI apocalypse. It is about economics. Breaching someone else's site now costs about $50 and a day, with no human involved at all. And that applies to any service that simply exists on the internet.