AWS says AI agents found and exploited a MONAI Label zero-day
By AgentRiot Editorial
The agents gained control of a university server, AWS says. A fix is in MONAI Label’s source tree, but the current PyPI release is still affected.

Amazon Threat Intelligence says AI agents found a previously unknown flaw in MONAI Label, an open-source tool for annotating medical images, and used it to take control of a university server during a September campaign. In its November 20 investigation (opens in new tab), AWS says the agents went from examining the application to writing files on the server in about 11 minutes. AWS assesses with high confidence that the vulnerability was unknown at the time. The CVE record (opens in new tab) rates it 9.8 out of 10.
What the operation records show
AWS discovered the campaign on September 10 because the operator had left the system running the agents publicly accessible without authentication. AWS could read its prompts and activity logs. Those records describe about 48 hours of activity starting September 8, targeting addresses associated with universities in East Asia outside mainland China.
The logs show people and agents playing different roles. AWS attributes the initial discovery to the agents. In a later session, a human operator gave the agents access into a second university’s network. AWS infers that the operator used the same flaw to get that access; it does not present the second incident as another independently observed discovery. AWS says it notified the software maintainers and worked with a government partner to notify the affected organizations. The public report does not identify the universities or establish whether patient data was accessed.
Other parts of the campaign failed. In one case, AWS counted 148 sequential actions across roughly 16 hours against a university system without the agents taking control. That contrast matters more than a blanket claim that agents can break into anything: the same campaign produced both a fast compromise and hours of unsuccessful attempts. Defenders need a way to contain the first case without treating every alert in the second as proof of a breach.
There’s a source fix, but no patched PyPI release
AWS says the maintainers released a fix. The maintainer advisory (opens in new tab) lists monailabel versions 0.8.5 and earlier as affected and names no patched package version. The change referenced by the CVE record is merged into the project’s default branch and is not in the 0.8.5 release tag. PyPI (opens in new tab) still lists 0.8.5 as the current version as of November 21. Updating from the package index alone doesn’t apply the fix.
The maintainers advise keeping MONAI Label off untrusted networks and running it under an unprivileged account with writes limited to its datastore. For a server that was already reachable from an untrusted network, removing that access is the immediate step, not the end of the investigation. Preserve logs and review unexpected changes and service-account activity before returning it to use. If you can’t verify that your deployment contains the source fix, keep the network restriction in place rather than relying on a version number or waiting for the next package release.

