Amazon says autonomous AI agents used by an actor suspected of having ties to China found a previously unknown flaw in MONAI Label and gained the highest-level access to a university server, with the ability to write files, in about 11 minutes.
The open-source application helps researchers label medical images. Cairn, the framework used in the operation, coordinates AI agents that pursue goals with little human direction. Amazon said it uncovered the campaign after finding a publicly accessible Cairn server that exposed the operator's prompts and the agents' activity.
Records showed that universities in East Asia outside mainland China were targeted over roughly 48 hours in September. The agents mapped how other software could interact with MONAI Label and found a flaw in its file-upload feature that let them write files outside the intended storage area. In a later session, they used the flaw to put a Python program on the server that opened a remote command connection.
They ran the program and gained the server's highest-level access. Amazon assesses with moderate confidence that the actor had links to China, citing the language in the prompts, timestamps and choice of targets. The project maintainers were notified and published a security advisory.
As of reporting on Nov. 20, the fix had been added to the project's code, but a patched software release had not yet been issued. Administrators were advised to restrict network access and permissions while checking exposed systems for signs of compromise.