An AI Agent Army Hacked 100+ Sites: What We Know
The headline reads like science fiction: per The Verge, a hacker used an army of AI agents — operating mostly autonomously — to break into more than 100 websites, with Reuters reporting victims including John Deere, SharkNinja and Cummins. Beneath the sci-fi framing is a story about tooling, and it is less mysterious than it looks.
The verified skeleton
- Security firm Norforge documented the campaign, discovering it through exposed work logs on Cairn, an open-source "general-purpose problem-solving engine" where users give agents a start and end point and the agents document their progress.
- The most active hacker laid out preferred intrusion techniques in fewer than 100 Chinese characters — a reminder that at current capability levels, the human's job is strategy; the agents do the grinding.
- Successful breaches took about 23 hours on average. Patience, not genius, is the exploit.
- Norforge found 20+ exposed Cairn instances likely running malicious campaigns; Cummins confirmed it is investigating.
Why this is less (and more) than "AI hacker"
Less: the agents did not "invent" the hacks. They industrialized known intrusion patterns — the same way earlier tooling industrialized scanning. A patient actor with a checklist and compute was always dangerous; now the checklist executes itself. More: the attack surface discovery was operational security failure on the attacker's side too — exposed work logs — which is how the campaign became public at all.
The pattern-recognition takeaway
Expect two waves of coverage: stories that inflate this into "AI destroys cybersecurity," and stories that dismiss it as repackaged automation. Both miss the middle, where the actual risk lives — every autonomous-agent tool that documents its own progress is also, when misconfigured, a public confession. We will keep applying this checklist to agent stories in explainers; the skepticism toolkit from our Yandex debunk applies here almost word for word.