A company may have dozens of secondary websites that no one checks daily: support pages, vendor portals, or old campaign sites. That is precisely the door an attacker managed to open on a massive scale using artificial intelligence agents that worked with almost no human supervision.
The AI-specialized cybersecurity firm Norforge detected that a single hacker had compromised more than 100 websites using a set of autonomous agents. Among the victims, as reported by Reuters, are companies such as John Deere, SharkNinja, and Cummins. According to an account by The Verge, Norforge has not publicly identified the targets.
The case is significant because it shows, with concrete figures, how much the time and effort required to execute an intrusion at scale have been reduced when the work is delegated to AI systems capable of operating almost independently.
An exposed panel in Cairn uncovered the operation
Norforge researchers tracked the attacks through work that had remained visible in Cairn, an open-source platform described as a general-purpose problem-solving engine.
The operation of Cairn is simple in appearance: the user defines a starting point and an end goal, and the artificial intelligence agents document their progress as they move toward that objective. Norforge warns that this same mechanism can be repurposed to coordinate cyberattacks step by step.
23 hours of autonomous work per successful intrusion
The identified attacker was the most active among all the exposed Cairn instances found by Norforge. Their victims, according to Reuters, were companies located in the United States, Canada, Australia, and Japan.
To direct their agents, the attacker wrote their preferred intrusion tactics in less than 100 characters in Chinese. According to data collected by Reuters, each successful intrusion took about 23 hours to complete. In other words, a brief instruction was enough to set in motion a full day of autonomous work that ended with a compromised website.
Dozens of other instances considered malicious
The case is not an isolated incident. Norforge located more than 20 additional instances of Cairn that it considers likely malicious and states it is working to have them removed.
The creator of the platform, who identifies with the username @le1xia0, stated earlier this month that Cairn was never intended as a hacking tool. According to this developer, the misuse of the platform depends on human decisions and not on a feature of the engine itself.
The response from John Deere, SharkNinja, and Cummins
The three companies identified by Reuters have offered different explanations, and it is important to distinguish them precisely.
A spokesperson for John Deere stated that there is no evidence that sensitive company information has been accessed. SharkNinja, for its part, explained that the breach affected an independent site containing product support information and assured that it has already managed to contain the threat.
In the case of Cummins, spokesperson Melinda Koski told The Verge that the company is investigating the report of a possible cyberattack against a website operated by one of its third-party vendors. In other words, the affected site was not directly owned by Cummins, but by a contracted third party.
The pattern: peripheral sites, not main corporate websites
The three cases share a common trait. At SharkNinja, the compromised site was an independent web dedicated to support. At Cummins, the page is managed by an external vendor. Neither company has mentioned any impact on its main corporate website.
That pattern indicates where these types of attacks are aimed: secondary pages, usually less monitored than the company's digital core. An AI agent that does not get tired can test hundreds of these sites with the same level of attention it would dedicate to just one.
What a company can do against these types of threats
Based on what happened with Cairn, several practical measures can be extracted for those who manage websites themselves or through third parties.
It is advisable to create a complete inventory of secondary sites. Support, product, or marketing campaign websites are precisely the ones that appeared compromised in the SharkNinja and Cummins cases.
It is also key to ask vendors who operates each site linked to the brand. In the case of Cummins, the investigation falls on a site managed by a third party, which underscores the importance of being clear about contractual responsibilities before an incident occurs.
Another recommendation is to check if there are any internal instances of AI tools accessible from the internet without adequate protection. Norforge located the attackers precisely through publicly exposed Cairn panels.
Finally, it is advisable to have a specific containment plan for each site. SharkNinja managed to isolate the threat on an independent website, which demonstrates the value of maintaining separate systems.
What this case means for users and professionals
For those working in cybersecurity, the most significant data point is the time: about 23 hours per successful intrusion, starting from an instruction of less than 100 characters. Relying on discretion or going unnoticed is no longer a reasonable defense when the attacker delegates operational work to autonomous agents.
For users of the mentioned brands, the situation is still developing. John Deere asserts there is no evidence of access to sensitive data, and SharkNinja limits the impact to a support site that has already been contained. Cummins, meanwhile, continues to investigate the actual scope of the incident on its third-party vendor's site.
Frequently asked questions
What is Cairn and why was it used in this attack?
Cairn is an open-source platform that functions as a general-purpose problem-solving engine. It allows setting a starting point and an end point while AI agents document their progress. According to Norforge, that same mechanism was used to coordinate computer intrusions.
Which companies were affected by this attack?
Reuters identified John Deere, SharkNinja, and Cummins among the victims. Norforge, the firm that detected the attack, has not made the identity of the compromised targets public.
How long did it take the attacker to compromise each website?
According to data collected by Reuters, each successful intrusion took approximately 23 hours to complete, starting from instructions of less than 100 Chinese characters written by the attacker.
Were the main websites of these companies at risk?
Not according to the available information. In the cases of SharkNinja and Cummins, the affected sites were secondary support pages or those managed by external vendors, not the main corporate websites.
What does the creator of Cairn say about this use of their platform?
The developer, identified as @le1xia0, stated that Cairn was never designed as a hacking tool and that misuse depends on human decisions, not on a feature of the engine itself.
Have the detected malicious instances been removed?
Norforge identified more than 20 instances of Cairn that it considers likely malicious and states it is working to have them removed.
