An unprecedented cyber hacking campaign uses autonomous AI agents via the Cairn platform to compromise more than 100 websites at a cost of just $50 per successful operation.
Investigative report reveals exploitation of Cairn platform in widespread attacks
Mechanism of autonomous agents and technical adaptability
Low financial cost and models used in the campaign
Responses of affected companies and limits of actual impact from breaches
FAQs
Investigative report reveals exploitation of Cairn platform in widespread attacks
An investigative report by Reuters, supported by technical research from cybersecurity firm Norforge, has documented a large-scale cyberhacking campaign in which an attacker heavily relied on autonomous AI agents to scan, discover vulnerabilities in, and breach websites. This activity was spotted by security researcher Snorre Fagerland, senior threat researcher at Norforge, while examining exposed internet installations of an open-source agent orchestration platform known as Cairn.
According to the researcher’s findings, a Chinese-speaking attacker used artificial intelligence tools to attempt intrusions into more than 1,200 websites distributed across the United States, Canada, Australia, and Japan, successfully breaching over 100 of them. Reuters independently examined this activity and identified several affected organizations, including major companies such as agricultural equipment maker Deere, home appliance brand SharkNinja, and engine manufacturer Cummins.
Mechanism of autonomous agents and technical adaptability
Digital forensics data indicated that the offensive operation actually began on September 25, with Fagerland noticing part of this activity in early October before later identifying additional malicious deployments of agent orchestration software. Consequently, the campaign itself did not launch on October 19; rather, the new development consists of the investigative report’s publication outlining the full scope of the attacks, their methods, and targeted entities.
The report stated that the attacker provided relatively concise initial instructions, while AI agents carried out extended and complex sequences of technical tasks. The agents searched for vulnerabilities, attempted to exploit them programmatically, and scanned associated systems. In one documented case, the agents failed to find a vulnerability in the primary target application and automatically shifted to scan its associated infrastructure until they ultimately found a weakly protected control interface.
This behavior illustrates the fundamental difference between traditional automated scanning tools and an AI agent capable of adapting its sequence of actions in response to failed attempts. However, the report does not prove that every breach occurred in complete isolation from human direction, as a human operator remained responsible for setting targets and providing the platform with initial instructions.
Low financial cost and models used in the campaign
The estimated financial cost of this cyber campaign was striking. Researcher Fagerland estimated that the attacker spent only about $5,000 on AI models throughout the entire campaign, equating to roughly $50 per successful breach. Technical records showed the offensive activity relied on a mix of DeepSeek models and Anthropic’s Claude Code tool.
The report emphasizes that naming these models identifies the technologies used in the campaign, but in no way implies that the developing companies authorized or knowingly supported this malicious activity. This research highlights how advanced AI agents can reduce the human effort required to launch repetitive technical attacks, particularly when an attacker targets a vast number of systems and lets software run continuously without real-time supervision.
Responses of affected companies and limits of actual impact from breaches
Reuters contacted several organizations named in the breach logs. Deere stated that it took necessary remediation steps and found no indication that sensitive information was accessed. SharkNinja indicated that the affected digital asset was a separate website hosting product support information and that the threat had been fully contained, while Cummins reported it was investigating a potential incident involving a third-party vendor’s website.
These official responses prove that the actual impact varied from one organization to another. A successful breach of a secondary website does not necessarily equate to compromising a company’s main corporate network, nor does unauthorized access automatically constitute proof of customer data theft. The primary physical limitation in this investigation remains Reuters’ inability to independently determine the volume of data that may have been extracted as a result of these breaches, despite the report providing unprecedented concrete evidence of deploying autonomous agents in repetitive offensive operations.
The significance of this investigative report lies in shifting the ongoing debate surrounding AI risks in cybersecurity from a theoretical framework to field incidents backed by documented figures. A single attacker scanning over 1,200 websites and breaching more than 100 of them for a total cost of no more than $5,000 demonstrates that autonomous agents are radically altering the economics of cyberattacks by lowering the cost per attack to about $50.
In turn, this development forces corporate and institutional cybersecurity teams to comprehensively review mechanisms for protecting auxiliary digital assets, product support sites, and third-party vendor platforms. The responses from companies like Deere, SharkNinja, and Cummins showed that offensive agents automatically search for the weakest links in an organization’s associated infrastructure when they fail to breach the main interface, necessitating the fortification of all subdomains and exposed internet control panels.
FAQs
Question: How many websites were targeted by the AI agent-driven hacking campaign?
Answer: The campaign attempted to infiltrate more than 1,200 websites in the United States, Canada, Australia, and Japan, successfully breaching over 100 of them.
Question: What was the cost per breach using AI agents in this campaign?
Answer: Researchers estimated the attacker’s total spending on operating AI models at about $5,000, equivalent to roughly $50 for each successfully breached website.
Question: What platform and technical models did the attacker exploit to carry out the attacks?
Answer: The attacker relied on the open-source Cairn agent orchestration platform, running a combination of DeepSeek models and Anthropic’s Claude Code tool to execute technical tasks.
Question: What did Deere, SharkNinja, and Cummins say about the impact of these attacks?
Answer: Deere confirmed no sensitive data was accessed, SharkNinja stated the threat was contained on a secondary product support site, and Cummins is investigating a potential incident on a third-party vendor site.
