AI agents working for attackers found a previously unknown weakness in medical-imaging software and later used it to take full control of a university server, Amazon disclosed on November 20. The company’s account traces the break-in to a September campaign using Cairn, an open-source framework that coordinates AI agents for security testing.
An exposed server revealed the campaign
Amazon’s threat intelligence team found the attackers’ Cairn server on September 10. It was accessible without authentication, exposing the operator’s prompts. The campaign had begun September 8 and ran for about 48 hours against universities in East Asia outside mainland China. Amazon assessed a China connection with moderate confidence; prompts were written in Simplified Chinese, and executed scripts used China Standard Time.
The operator gave the agents a short goal: obtain shell access—the ability to run commands on a target machine. Success was not immediate. According to Amazon CISO CJ Moses, the agents spent about 16 hours taking 148 sequential actions against another university’s Redis server. They tried several routes, including loading unauthorized code and changing configuration files, but none succeeded. They also attempted attacks against other services and tried multiple passwords across accounts.
A separate operation in the campaign spent more than six hours targeting a university GPU server. The agents assembled 80 real usernames from university websites, academic papers and other public documents, then used them in password-guessing attempts. That operation also failed, Moses said.
Neither operation succeeded, yet together they show that the agents kept adapting for nearly a full day without human direction
CJ Moses, Amazon CISO
A file upload became a route to control
The successful route involved MONAI Label, an open-source medical-imaging AI application. The agents identified its software framework, mapped the interfaces used to communicate with it and found an upload function with a caller-controlled filename. Amazon says they went from mapping those interfaces to writing files outside the intended storage directory in about 11 minutes.
The flaw is called path traversal: supplied file paths can escape the folder meant to contain them. MONAI Label’s image and label upload functions built destinations from user-controlled values without checking that the resulting paths stayed inside the datastore. The project’s advisory says attackers could write anywhere the application had permission to write, potentially overwriting application code or configuration files.
The 11-minute interval was not the time to full takeover. In a later session, the agents used the same flaw to run a Python program that gave them remote command access as root, the server’s highest-privilege account. Human involvement remained visible: in another session, the operator supplied a proxy into a second university network containing the same vulnerable code.
Amazon assessed with high confidence that this was a zero-day—a previously unknown vulnerability. It found no earlier fix, issue, advisory or vulnerability-database mention at the time of discovery. That assessment concerns the flaw’s prior disclosure, separate from the company’s less certain attribution of the campaign.
The fix has not reached a release
Amazon notified the maintainers. A fix was merged into the project’s main source-code branch on October 19, and the critical security advisory followed on November 18. It covers version 0.8.5 and earlier and lists no patched version. VentureBeat reported on November 20 that 0.8.5 remained the newest release, leaving the source-code repair distinct from an available packaged update.
The repair rejects destinations outside the permitted image and label directories and adds checks on image identifiers. Until a release includes it, exposure depends partly on deployment settings. Authentication is disabled by default, allowing any reachable network client to exploit the flaw. Enabling it narrows access, but users with the ANNOTATOR role can still exploit the vulnerability.
The advisory’s interim safeguards
- Enable authentication and grant the ANNOTATOR role only to trusted users.
- Keep MONAI Label servers off untrusted networks.
- Run the server without administrator privileges and limit write access to its datastore directory.
Administrators should also check for signs of an earlier intrusion, including unexpected changes to application files, VentureBeat advised. Tightening access now does not undo a compromise that has already happened.
Reader comments
Newest comments first. Replies stay oldest first.