A file upload could escape its intended directory

In research published on 20 November, AWS described observing an exposed Cairn instance used by an actor it assesses as Chinese-nexus with moderate confidence. The agents found a path traversal flaw in MONAI Label's datastore upload endpoints, then used it to write outside the intended datastore. In a later session, they wrote a Python payload to sitecustomize.py and triggered a process that ran it.

The maintainer advisory confirms that versions through 0.8.5 are affected. Authentication is off by default; if the server is reachable in that configuration, a network client can reach the vulnerable endpoints. With authentication enabled, an untrusted user granted the ANNOTATOR role still has the necessary access.

A merged fix is not a released package

The maintainer has merged a code fix, but the advisory currently lists no patched version. Do not mark a deployment fixed merely because a pull request has merged; check the version actually installed and watch for a release.

Reduce reach and write access now

  1. 1

    Locate MONAI Label servers. Find deployed versions and determine whether the datastore image or label upload endpoints are reachable from untrusted networks.

  2. 2

    Use the maintainer’s workarounds. Enable authentication, avoid granting ANNOTATOR to untrusted users, restrict network access and run the service without elevated privileges. Limit its write access to the datastore.

  3. 3

    Look for file-write consequences. Review upload requests and unexpected files outside the datastore, especially Python startup files, application code and scheduled-task locations. Investigate new processes and outbound connections from the MONAI Label host.

The service account decides the damage

Path traversal lets a file land somewhere its caller did not intend; what happens next depends on what the service process can write and execute. The AI-agent detail explains the observed discovery speed, but exposure, authentication and filesystem permissions determine the local risk. AWS's attribution is an assessment about the observed operator, not every use of Cairn.