Science & Tech
Autonomous AI agents spend 23 hours breaching corporate websites
Persistent machine-led attacks challenge defenses built for human intruders
Persistent machine-led attacks challenge defenses built for human intruders
A Chinese-speaking hacker used to probe and breach corporate websites with little human oversight, starting September 25 across the United States, Canada, Australia and Japan. The agents searched for exposed systems, tested weaknesses and adapted their attacks, attempting to enter more than 1,200 sites and reportedly compromising more than 100. Successful intrusions took about 23 hours, while the campaign cost roughly $5,000 in AI usage, making large-scale attacks relatively cheap. researcher Snorre Fagerland identified the activity while monitoring exposed and linked it to and .1
The impact varied. Deere said it remediated its website and found no evidence of sensitive-data access. SharkNinja said a contained incident affected a separate product-support site, while Cummins investigated a possible attack on a . The episode shows how defenses may miss persistent, automated activity spread across company and supplier systems. It also raises unresolved questions about the true number of victims, what data was accessed and whether security teams can monitor AI agents fast enough without stronger human oversight.1
Reports described repeatedly probing websites for extended periods with limited human direction, highlighting challenges for defenders. The full victim count, data access and attribution remained unresolved.
Cummins investigated activity involving a website operated by an outside vendor. The case showed how outsourced web infrastructure can complicate visibility, ownership and response when security teams cannot directly control every affected asset.
SharkNinja said the affected site was a containing support information. The company said the threat had been , indicating that separation from other systems limited the reported impact.
Deere confirmed that its website was among the affected targets and said it took necessary measures. The company reported no evidence that sensitive information had been accessed, distinguishing the website intrusion from confirmed data theft.
The reported successful attacks involved prolonged testing and repeated attempts to find exploitable weaknesses. Fagerland estimated that the campaign consumed roughly $5,000 in , or about $50 per successful intrusion.
By late the following week, the campaign had attempted intrusions against more than 1,200 sites across multiple industries. More than 100 company websites were reportedly compromised, although attempted access and were not equivalent findings.
Fagerland observed the activity while investigating internet-exposed instances of , an open-source AI problem-solving platform. He described the operator as Chinese-speaking and assessed that the campaign used a combination of and Anthropic's .
The agents searched for exposed systems, examined potential weaknesses and adapted their next steps based on what they found. This differed from an AI tool merely suggesting commands, because the agents pursued a multi-step attack workflow with little direct intervention.
A cybersecurity researcher reported that, beginning September 25, 2026, a Chinese-speaking hacker directed at e-commerce websites in the United States, Canada, Australia and Japan. The agents attempted to probe more than 1,200 sites, with reported successful intrusions involving more than 100 companies and limited human oversight.