A campaign run mostly by machines

On Monday, 19 October, Snorre Fagerland, principal threat researcher at NorForge, published findings describing a Chinese-speaking hacker who breached more than 100 company websites while doing very little of the work personally. Most of the operational effort fell to autonomous AI agents, which were set in motion by instructions of fewer than 100 Chinese characters, according to Reuters. The agents searched for weaknesses, tested ways to exploit them and gained access with little human input. The tool behind the campaign was Cairn, an open-source AI platform that takes a target and a goal and assigns agents to pursue it in stages. Agents post what they find to a shared board that every agent can read, and a human operator can add hints along the way.

How the exposed panels were found

Many Cairn operators left their control panels reachable on the open web, and Fagerland located them using internet scanning tools. He began spotting them in early October, while the hacker's campaign had started on 25 September. He judged more than 20 panels likely malicious, and most of the activity fell into three clusters. The Chinese-language cluster, which NorForge rates as malicious with high confidence, was the busiest, with more than 1,100 attack projects spread across 14 panels. Its typical task targeted online shops in the United States, Canada, Japan and Australia, with traffic levels of 200,000 to 1.5 million. An automated pipeline supplied those tasks, instructing agents to find a route to run code on each site and then install an uploader web shell, a form that saves uploaded files rather than running commands. NorForge believes the goal was lasting access to shop and payment systems.

Victims and their responses

Reuters said it identified some victims by reviewing the hacker's activity on Cairn, naming farm equipment maker Deere, appliance company SharkNinja and engine maker Cummins. Deere told Reuters it carried out remediation and saw no sign that anyone reached sensitive information. SharkNinja said the intrusion affected a standalone product support site and that the threat had been contained. Cummins said it was examining a report of a possible attack on a website run by one of its vendors. Reuters could not establish how much data, if any, was taken, so the scale of any data loss remains unknown.

Cost and speed of the attacks

Reuters put the cost of a successful intrusion at about $50 and the entire campaign at about $5,000. A successful attack took the agents about 23 hours on average. The agents tried more than 1,200 websites in total, and the campaign reached more than 100 companies. The gap between those two figures shows how cheap automated trial and error can be when no one has to supervise each step.

Cost per successful intrusion50

Total campaign cost5,000

Cost per successful intrusion

50 USD

Total campaign cost

5000 USD

Data from the article text

Wider warnings about AI-driven cybercrime

The Israeli firm Gambit Security reported a separate case in which Russian-speaking hackers used AI tools in August to breach seven organisations. In September, similar attacks on 25 firms led to the theft of data from at least 600,000 payment cards, according to the same firm. Experts cited in the coverage warn that autonomous AI tools are sharply lowering the cost of cybercrime. NorForge said it is working to have the malicious Cairn instances taken down. The creator of Cairn, who posts as @le1xia0, has said the platform is not a hacking tool and that misuse is a human decision.