Cairn AI agents found an unknown flaw in a medical imaging application and used it to take full control of a server, with the attackers giving them only a short instruction. Amazon Threat Intelligence says the suspected Chinese-linked operators told the agents to “obtain shell access” and left them to work out the rest.
Cairn is a legitimate penetration-testing tool. It coordinates several AI agents and was built for research and defense, which makes its misuse a concern for security teams.
How the campaign came to light
Security teams have worried for a while that agentic tools would lower the bar for attackers. This case shows what that looks like in practice: no custom exploit code, just a goal stated in plain language and software that kept trying until something worked.
Amazon’s researchers found the operation because the attackers left their Cairn instance publicly reachable with no authentication. That exposed the prompts, targets and logs of the whole effort.
The activity ran between September 8 and 10, 2026, and mostly targeted educational institutions in East Asia outside China. Amazon attributes it to a Chinese-nexus threat actor with moderate confidence. The assessment rests on the language in the recovered prompts, the timestamps of the activity and the choice of victims.

From API mapping to root in minutes
The agents found a path traversal zero-day in MONAI Label, an open-source application that uses AI to help researchers label medical images. The bug let them write files to arbitrary locations on the server.
It took the agents 11 minutes to go from mapping the application’s API to exploiting the flaw. In a later session they wrote a Python reverse shell into the sitecustomize.py file, which loads automatically whenever Python starts. They then called an API endpoint to trigger it, and that gave them root access.
The operator then supplied a proxy into a second university network running the same vulnerable software and told the agents to go after other machines there.

Where the agents failed
Cairn AI agents still hit walls, and the failures are as telling as the success.
The agents were not unstoppable. One effort to take over a Redis server ran for 16 hours and 148 actions and failed. Against a university GPU server, they tried more than 30,000 credential combinations over six hours without getting in.
The record is mixed. Brute force went nowhere, yet a single open-ended instruction was enough to produce a working exploit chain against an unpatched program.
Not the first Cairn abuse
Other security teams have seen the tool turned against victims. Gambit Security documented Cairn abuse in September as part of a large credit card theft operation. More recently, Google reported its use in an attack on a U.S. aerospace organization.
Amazon published two IP addresses tied to the campaign, both seen between September 8 and 10, 2026. Defenders can search their logs for them.
What defenders should do now
MONAI Label is popular with academic labs, which often run research software with few of the controls a company would apply. That is why a campaign aimed at universities fits the pattern.
- Update MONAI Label. Amazon reported the flaw to the maintainers, who released a security update on November 18.
- Keep any agent or testing platform behind authentication. The attackers’ own unprotected Cairn server is how they were caught.
- Review who can reach research servers, including GPU machines, from the open internet.
What happens next
Amazon’s warning is blunt: AI agents can find and exploit vulnerabilities faster than security teams can respond. It urges organizations to lean more on automated defenses instead of manual investigation.
Expect more reports like this one, because the tool is cheap, scriptable and easy to point at any internet-facing service. Cairn AI agents were already linked to several incidents within a few weeks, and each one lowers the skill needed for an attack. Universities and research labs running open-source tools with default settings are likely to be early targets. For them, quick patching matters more than ever.
Source: BleepingComputer