MONAI Label flaw linked to AI-agent intrusion at a university

Amazon Threat Intelligence says autonomous AI agents exploited a previously unknown vulnerability in MONAI Label, an open-source medical-imaging server, and obtained root access on one university server. The issue affects MONAI Label 0.8.5 and earlier in the default configuration, where authentication is disabled. It is now tracked as CVE-2026-133968 with a critical CVSS score of 9.8 out of 10.
MONAI Label is used by research teams to label medical scans and train imaging models. It is separate from the broader MONAI framework. The project maintainers merged a source-code correction on October 19, but its advisory says a corrected release is still due in a future version. The newest listed release remains version 0.8.5 from 2024.
An unauthenticated write outside the storage folder
The vulnerability is a path-traversal weakness in components that save uploaded images and labels. A caller-controlled file identifier was appended to the storage directory and written to disk without ensuring that the resulting path stayed inside the approved location.
That condition allows an unauthenticated client able to reach the server to write files beyond MONAI Label's intended storage folder. The practical impact depends on the privileges of the account running the service: the flaw does not automatically permit writing anywhere on the host, but a writable code or configuration location can create a route to code execution.
The merged fix resolves upload paths within approved folders and rejects empty names and names beginning with a dot. Organisations using a source build should confirm that this specific correction is present, because a merge in the repository is not equivalent to an installable patched package.
Amazon reconstruction shows agent-led reconnaissance
Amazon reconstructed the activity after discovering an internet-exposed Cairn control panel without a password. Cairn is an open-source framework for running AI agents as automated penetration testers. Its exposed settings, prompts and run history showed that the operator chose targets and set a goal of obtaining shell access, while the agents selected methods with little step-by-step direction.
Over roughly 48 hours in September, Amazon said the agents mapped the application interface and reached the ability to write files in about 11 minutes. The later root-shell result used the same flaw to place a Python startup file, then triggered a training feature that launched a new Python process. That process loaded the planted file and connected back to the attacker as root.
The result was observed on one university server, not established as universal across MONAI Label deployments. A second university network was later accessed through a tunnel supplied to the agents; Amazon believes it was compromised in the same manner but has not confirmed that conclusion. Investigators assessed a China-linked actor with moderate confidence based on Simplified Chinese prompts, China Standard Time settings and target selection, without attributing the activity to a named group or government.
Contain exposure while a release is pending
This incident also reinforces the operational risk highlighted by AI discovery of 21 FFmpeg zero-days as automated systems accelerate the discovery and testing of zero-day weaknesses. The reported campaign included failed operations, and Amazon's account has not been independently confirmed; it does not establish that patient data was taken, clinical systems were accessed, or other deployments produced root access.
Until a patched version is released, teams should block untrusted network access with firewall and routing controls, enable authentication, and run MONAI Label under a restricted account with tightly limited writable folders. Administrators should also review any reachable deployment for unexpected Python startup files and unusual training jobs. The immediate business implication is to treat exposed, unauthenticated research services as a priority containment task rather than waiting for a package update.

