MONAI Label 0.8.6 Fixes Exploited Flaw: What Server Operators Should Do

MONAI Label 0.8.6 fixes an exploited file-write flaw. Check exposed servers for signs of intrusion, upgrade and restrict access to the imaging service.
Server racks in a data center used to illustrate enterprise networking and security systems
Photo by Kevin Ache on Unsplash

Project MONAI has released MONAI Label 0.8.6 to fix a critical file-write vulnerability in its medical-imaging annotation server. The update follows Amazon Threat Intelligence’s report that an attacker used AI agents to discover and exploit a flaw in an open-source medical imaging application during a September campaign against universities.

The November 22 release identifies the datastore vulnerability as CVE-2026-133968 and urges all MONAI Label server deployments to upgrade. Versions through 0.8.5 are affected. Administrators should also investigate older servers that were reachable from untrusted networks: installing a fix does not remove an attacker who is already inside.

Amazon’s incident report links to the MONAI Label advisory but does not name the application in its narrative. It describes one observed compromise, not evidence that every exposed MONAI deployment was attacked or that patient records were stolen.

How the MONAI Label flaw works

MONAI Label helps researchers and clinicians annotate medical images for machine-learning workflows. Its server provides upload and training interfaces that can be reached over a network.

According to the project’s security advisory, the image-upload endpoint accepted a caller-controlled image identifier without checking that the resulting file path stayed inside the datastore. A related label-upload operation had the same problem. A network client could therefore write files wherever the server process had permission, potentially leading to code execution or damaged configuration.

Authentication is disabled by default. When it is enabled, a user with the ANNOTATOR role can still reach the affected functionality, so authentication alone is not a substitute for updating or restricting trusted users.

Upgrade, then verify the running deployment

The project lists version 0.8.6 as a security release and provides these update commands for its Python package and container image:

pip install -U monailabel==0.8.6
docker pull projectmonai/monailabel:0.8.6

Run the package command in the environment actually used by the server. Pulling a container image does not update an existing running container: recreate or redeploy the service with the updated image under your normal change process, then verify the running version. Check research workstations, managed servers and container deployments rather than assuming a single installation represents the whole environment.

Plan for compatibility before rollout: the release requires Python 3.11 or newer and removes support for Python 3.9 and 3.10. The project also notes changes to its image and dependencies. Review the release notes and test integrations where practical, while prioritizing exposed systems.

Version 0.8.6 also fixes a separate session-upload path traversal involving archive extraction and session IDs. The project identifies both issues as affecting versions through 0.8.5; Amazon’s report should not be read as evidence that the separate session flaw was used in its observed incident.

Check for intrusion and reduce exposure

The maintainers recommend reviewing servers running 0.8.5 or earlier that were accessible from untrusted networks for unexpected file changes before returning them to service. Preserve logs and other evidence and follow your incident-response process if compromise is suspected.

  • Review unexpected changes to application files, Python startup files such as sitecustomize.py, SSH authorized keys and scheduled tasks, especially where the service account could write.
  • Inspect available server and reverse-proxy logs for suspicious requests to PUT /datastore/image and PUT /datastore/label, including path traversal attempts. Logs may not retain all request details, and an absence of matches does not prove a server was untouched.
  • Keep the server off untrusted networks, enable authentication, grant roles only to trusted users and run the service as an unprivileged account with write access limited to the datastore. These controls remain useful after upgrading.

Amazon published two network indicators in its report, observed in records from September 8 to 10. They can help a retrospective investigation, but they are not a complete list of attacker infrastructure. Treat a confirmed compromise as a broader incident, including assessing credentials and other systems the server could reach.

What Amazon observed

Amazon found an internet-accessible instance of Cairn, an open-source framework that coordinates AI agents, with its operational history exposed. The operator supplied targets and a goal of obtaining shell access; the agents planned and adjusted their steps within that task, according to the report.

In the successful operation, the agents went from mapping the application’s API to an arbitrary file write in about 11 minutes. In a later session they used the vulnerability to place a Python startup file containing a reverse shell, then triggered a new Python process. The resulting shell ran as root because of the permissions on that particular server. The 11-minute figure describes the path to the file write, not the full time to root access.

Other observed operations failed, including a lengthy attempt against a Redis server and a separate credential-spraying operation. Amazon assessed the actor as China-linked with moderate confidence, based on factors including the prompts, timestamps and targets. Its report also describes a later human intervention involving a proxy into another university network.

Leave a Reply

Your email address will not be published. Required fields are marked *

Previous Post
Server racks in a data center used to illustrate enterprise networking and security systems

VMware vCenter Zero-Day: How to Patch CVE-2026-123890 and Check for Intrusions

Related Posts