Norforge detects traces of AI-driven cyber operations on exposed control panels
October 22, 2026 · 2 min read
Also published in Português (Brasil)
Security firm Norforge announced that it found traces of malicious cyber operations conducted by artificial intelligence agents on more than 20 exposed control panels. The research, dated October 19, indicates that attackers assigned targets to these agents and left the execution of the operations largely to the software. The central issue in this incident is the use of agents to serve the objectives of an external operator, rather than the unauthorized actions of the AI systems themselves.
Norforge examined Cairn panels that were accessible to the internet without protection and identified distinct clusters of activity. The company reported that the most active cluster contained over 1,100 attack projects across 14 panels. Researchers noted that the number of projects does not equate to the number of successful attacks. They specifically highlighted that targets, task definitions, and success criteria varied significantly between these different clusters.
The report includes an approximate 10 percent success rate, which Norforge clarifies is not a universal test result applicable to all AI tools. This figure represents an estimate derived from the specific operations observed. The company also acknowledged that its cost estimates are not based on precise token consumption. These distinctions are important to prevent the report from being misinterpreted as a general security metric for how easily all websites can be compromised.
The investigation also revealed security flaws within the attackers' own infrastructure. Unprotected panels exposed findings related to targets, task history, and sensitive access information. Norforge stated that some teams conducting what appeared to be legitimate security tests faced the same issue. The lack of protection for the infrastructure running a test can create a secondary data leak risk for the organization being tested.
Public responses from affected companies limit the scope of the incident. John Deere, SharkNinja, and Cummins were named in the report. A John Deere spokesperson said there were no signs of access to sensitive information. Cummins stated it was investigating a potential attack on a site operated by a third-party supplier, which does not imply a breach of its entire internal systems. SharkNinja explained that the incident affected an independent site containing product support information and that the threat has been contained. Access to a website, exfiltration of sensitive data, and intrusion into a corporate network represent different outcomes, and public statements do not confirm the same scope for each company.
Norforge said it informed the affected organizations and worked to have the relevant panels closed. The company is not publishing the full addresses of some panels because they remain live and contain sensitive information. The report emphasizes that the security issue is not only about agent capabilities but also about how teams using these tools protect their control panels and collected customer data.