Content on Harava News is AI-generated and may contain errors.

How Did AI Hack MONAI So Fast?

FocusCrime

The news behind this article: Amazon Says Suspected China-Linked Actor's AI Agents Exploited MONAI Label Flaw in 11 Minutes

Amazon says autonomous agents found a previously unknown flaw in medical-imaging software and later gained root access to a university server.

A flaw in file uploads

MONAI Label is an open-source server that helps researchers and clinicians mark up medical images and build AI models. Its image-upload function used a name supplied by the caller to choose where to save a file, without checking that the destination stayed inside the intended folder. A remote user could therefore write files elsewhere on the server.

Amazon says its Cairn agents identified the application, mapped its API and tested the upload function. They reached arbitrary file writing in about 11 minutes. That was not the time it took to gain root access: in a later, separate session, they placed code where Python would load it and triggered another process. The server process was running as root.

Fast, but not unstoppable

The agents had a clear target and a simple goal: obtain shell access. They could keep testing and adjusting without an operator choosing every next move. Cairn coordinates AI agents to explore possible routes to a goal. The underlying weakness, however, was a familiar failure to restrict file paths, not a new kind of vulnerability created by AI.

Amazon also describes long attempts against other university systems that failed, including a 16-hour effort against Redis and thousands of unsuccessful login guesses. The campaign targeted university-linked IP addresses in East Asia outside mainland China for roughly 48 hours, beginning September 8, 2026. Amazon found an unauthenticated Cairn server on September 10. It assesses the actor as China-nexus with moderate confidence, citing prompts, timestamps and target choices, but has not identified the group or established a motive.

What is known and what comes next

The project’s advisory and code changes make the flaw and repair inspectable. The Cloud Security Alliance’s technical note explains the flaw, but its account of the September campaign draws on Amazon’s. Amazon’s public report does not include Cairn’s underlying logs or name the universities. The timing, agents’ role and attribution remain Amazon’s reported findings, not a sequence independently reproduced from public evidence.

Possible targets include research images, datasets, GPU resources and other university systems. The advisory covers versions through 0.8.5. Authentication is off by default, and the CSA says even an authenticated user with the ANNOTATOR role could exploit the flaw. No public reporting so far says clinical systems or patient data were affected.

The fix was merged into the project’s code in October. A November 22 report found version 0.8.6 tagged on GitHub, while PyPI still listed 0.8.5. The fixed version needs to reach normal distribution and deployed servers. Operators should also check for signs of compromise, because installing a fix cannot undo an intrusion that already happened.