Skip to content
Security & Trust

Amazon: AI agents exploited a MONAI Label file-path flaw 11 minutes after mapping its API

In one campaign, AI agents failed against Redis and a GPU server, then broke a medical imaging server through an unchecked file path

W
WebPulse Newsroom
AI-assisted · 4 min read
Share on X LinkedIn
Amazon: AI agents exploited a MONAI Label file-path flaw 11 minutes after mapping its API

Graphic: WebPulse, drawn from the sourced facts in this story.

In brief
  • Amazon says AI agents told only to 'obtain shell access' exploited a MONAI Label file-path flaw 11 minutes after mapping its API.
  • The same campaign failed against a Redis server and a university GPU server, so the agents did not succeed everywhere.
  • Teams running MONAI Label should turn on authentication, confirm the fix is in their build, and limit what the server process can write.

A common mistake met a patient machine

Some servers fall to a simple error that nobody checked. Amazon Threat Intelligence describes one such case. AI agents took over a medical imaging server because it trusted a file name it was given.

Amazon investigated a campaign run on Cairn, an AI agent platform. BleepingComputer reported the findings. Amazon says the operator told the agents only to "obtain shell access". The agents worked out the rest.

Two failures and one success

The campaign did not win everywhere. A Redis server held out against a 16-hour effort of 148 actions. A university GPU server also held. The agents tried more than 30,000 password pairs over six hours and got nowhere.

MONAI Label was different. It is an open-source tool that helps researchers label medical images. Amazon says the agents discovered a previously unknown path traversal flaw in it.

Amazon says 11 minutes passed between mapping the tool's API and exploiting the flaw. That figure covers only this step. Root access came in a later session.

How the flaw works

The maintainers' advisory explains the bug. MONAI Label accepts image uploads at a PUT /datastore/image endpoint. The caller supplies an image name in a query parameter. The server uses that name as the file's ID without checking it.

The server adds a file extension and joins the result onto the datastore folder. It then saves the upload. It never confirms that the final path is still inside that folder.

A name such as ../../../tmp/x climbs out of the folder. The file then lands anywhere the server process can write. The label upload endpoint builds its path the same way.

The advisory says a write like this can turn into remote code execution. An attacker could overwrite the application's own code. Another route is to plant a login key in .ssh/authorized_keys, or a job in cron.

Amazon says the agents chose a different route. They wrote a Python reverse shell into sitecustomize.py. A reverse shell makes the server call back to the attacker. Python loads that file on its own at startup. The agents then triggered an API endpoint and gained root access.

A default setting widened the door

The advisory says authentication is off by default (MONAI_LABEL_AUTH_ENABLE=False). With that default, any network client that can reach the server can use the flaw. With authentication on, the ANNOTATOR role is enough.

Amazon says the operator later supplied a proxy into a second university network. That network ran the same software. The operator told the agents to target other machines there.

Amazon attributes the campaign, with moderate confidence, to a Chinese-nexus actor. It says the attacks ran from September 8 to 10, 2026. They mainly hit East Asian educational institutions outside China.

Amazon learned all this because the attackers left their own Cairn instance open to the public without authentication.

What this case argues

The limits come first. This is one campaign, described by one investigator. The agents failed in two documented cases. Cairn is also a legitimate pentesting tool. BleepingComputer notes that Gambit Security and Google reported separate abuse of it. None of this shows a trend.

Within those limits, the case shows something useful. A one-line goal led the agents to a working attack on this server. The operator still stepped in later to direct further targeting.

The flaw belongs to a common class: a file path that was never checked. Still, the agents had to find it on their own. They then had to chain it into a shell.

Our argument is narrower than a trend. A reachable server with a plain flaw is the kind of target such agents can use. This one case does not show that every such server would fall. Two of the targets held.

Amazon's own warning is that AI agents can find and exploit flaws faster than security teams can react. It urges more reliance on automated defenses than on manual investigation.

Questions to put to your team

Servers run by research groups can sit outside the central security inventory. Check whether yours do. The advisory and BleepingComputer describe MONAI Label as a tool for researchers.

First, do we run MONAI Label, and is authentication switched on? The maintainers advise turning it on. They also advise keeping the ANNOTATOR role from untrusted users. They say to keep the server off untrusted networks.

Second, which version do we run? BleepingComputer says maintainers released a fix on November 18. The advisory says the fix is merged and will ship in the next release. Confirm that your build includes it.

Third, what account runs the server, and where can it write? The advisory advises a low-privilege account that can write only to the datastore folder.

Fourth, would anyone notice an unexpected change to a startup file such as sitecustomize.py? If the answer depends on a person looking, Amazon's advice points toward automation.

The agents failed against some targets. They succeeded against a server that trusted a file name. Defenders can ask which of their own systems do the same.

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Project-MONAI (MONAI Label maintainers).