Summary
NorForge reports finding more than 20 publicly exposed Cairn control panels likely linked to malicious activity. The panels reveal how operators used AI agents to probe websites, pursue access and, in at least one case, leave an uploader web shell.
NorForge says an investigation of publicly accessible Cairn control panels found more than 20 likely associated with malicious cyber activity. The panels exposed project records showing how operators tasked AI agents with probing websites for vulnerabilities and, in some cases, pursuing persistent access. NorForge also reports finding exposed credentials and evidence of a web shell on a targeted site.
Cairn is a state-based problem-solving system that can coordinate agents towards a goal set by an operator. In the cases NorForge examined, tasks included searching for weaknesses such as insecure file uploads, authentication flaws and routes to remote code execution. The investigation offers a view of how these tools were being used in intrusion attempts—and how poorly secured control panels could put both targets and operators at risk.
How Cairn organised the intrusion tasks
Cairn records each job as a project, with an initial state, a goal and a sequence of steps. Agents add discovered information, or “facts”, to a shared project board and post proposed actions, or “intents”. Human operators can add their own hints to the same board, so the work described by NorForge was not simply an agent acting without direction.
The largest activity cluster NorForge identified had 14 associated dashboards and more than 1,100 attack projects. Its targets were mainly organisations in the United States, Japan and Australia, with a focus on e-commerce and payment infrastructure. A second cluster had 10 panels and tasks oriented towards financial theft. A third, represented by one panel with 59 projects, was assessed as malicious with low confidence.
The task records included instructions to find unauthenticated ways into web applications and to install uploader shells. NorForge says it checked a shell URL listed in a completed task against a large hardware-parts vendor and found an uploader web shell serving from the site. The report describes this as a way for attackers to retain access; the shell could later be changed to carry out more harmful actions.
NorForge estimates that around 10% of tasks reached their stated goals. It cautions that the projects were not directly comparable: operators set different goals, and some tasks that did not reach a final goal still gathered information or achieved partial results. For projects in the first cluster that did meet their goals, NorForge reports an average of just under 23 hours from start to completion and about 140 facts per project.
Exposed control panels put more than targets at risk
The dashboards contained details about targets, agent discoveries and, in some cases, credentials. NorForge says its scan found hundreds of secrets of different kinds, with 11% verified as valid. That exposure can create risk even when a penetration test does not ultimately compromise its target, because the panel may also reveal information gathered during the work.
The researchers also found interfaces that appeared to expose controls for starting or stopping projects, changing their inputs and redirecting agents. They did not test whether those controls could be used without authorisation. NorForge withheld panel addresses because some were still live and contained sensitive information.
The report’s central warning is about operational security as much as AI capability: an agent-assisted attack can combine automated reconnaissance with human direction, while an unsecured dashboard can expose the work and secrets to outsiders. The investigation documents particular campaigns and panels; its success-rate estimate is specific to the varied tasks NorForge observed.