Cairn AI agents discovered a previously unknown flaw in medical imaging software during a September campaign that mainly targeted educational institutions in East Asia. According to Amazon Threat Intelligence, the agents took 11 minutes to find and exploit the flaw. They gained full control of the server in a later session.

An exposed dashboard revealed the operators’ instructions

Amazon researchers uncovered the campaign because the operators left their Cairn dashboard open to the public without a password. Its records showed what the operators asked the agents to do and which decisions the software made on its own. Cairn is a legitimate security testing platform that lets researchers coordinate several AI agents.

In one session, an operator gave the agents a goal: gain command-line access to a server. The instruction left the choice of method to the software. The agents discovered a flaw in MONAI Label, an open-source application that helps researchers annotate medical images. The flaw let them write files outside the locations the application normally permits.

The 11-minute interval began when the agents started examining the application’s interface and ended when they demonstrated they could exploit the flaw. In a subsequent session, the agents used the same vulnerability to gain administrator privileges. An operator then gave them access to a second university network running MONAI Label and directed them toward other machines. Amazon reported the vulnerability to the project’s maintainers, who released a security update on November 18.

Two other targets resisted hours of automated attempts

The campaign records also captured two failures. In one case, the Cairn AI agents spent 16 hours trying to take control of a Redis server. They performed 148 actions without gaining control. At another university server, the agents tried more than 30,000 username and password combinations over six hours. None gave them access.

Amazon traced the activity to September 8–10, 2026. The targets were mainly educational institutions in East Asia outside China. Researchers assessed a possible link to China with moderate confidence, citing the language of the operators’ instructions, the timing of the activity and the choice of targets.

For defenders, Amazon highlighted the short interval between the agents’ first look at MONAI Label and their successful use of the flaw. The researchers urged organizations to automate more of their defenses because a manual investigation can take longer than that initial attack.


0 responses to “Cairn AI agents find new flaw and take over imaging server”