Suspected China-linked hackers used autonomous AI agents to discover a previously unknown flaw in MONAI Label, an open-source medical imaging server, and subsequently gained root access to a university server, according to Amazon Threat Intelligence.
The vulnerability, now tracked as CVE-2026-133968 with a critical score of 9.8, permits an unauthenticated attacker to write files outside the server’s designated storage folder. It affects MONAI Label 0.8.5 and earlier versions when running in the default configuration that leaves authentication disabled. The server is widely used by research teams to label medical scans and train imaging models.
The weakness is a classic path-traversal issue: a caller-supplied file identifier is concatenated with the storage directory and written to disk without validation that the resulting path remains inside the intended folder. Because the server executes with the privileges of its runtime account, an attacker who can reach the service over the network can place files in writable locations that may later lead to code execution.
Attack Timeline and AI Agent Operations
Amazon reconstructed the intrusion after locating the operator’s Cairn control panel exposed on the internet without authentication. Cairn is an open-source framework that orchestrates AI agents as automated penetration testers. The exposed panel disclosed task settings, prompts, and run history showing that the agents worked with minimal step-by-step guidance.
Over approximately 48 hours in September the agents mapped the application interface and reached the stage of writing arbitrary files within roughly 11 minutes. The operator defined the high-level goal of obtaining shell access and selected targets; the agents independently chose the exploitation methods. In a later session the same agents used the flaw to drop a file that Python automatically executes at startup, then triggered a training feature that loaded the file and established a reverse shell running as root.
Investigators linked the campaign to a China-linked actor with moderate confidence, citing Simplified Chinese prompts, China Standard Time settings, and target selection consistent with Chinese interests. No named group or direct government attribution was established.
Current Status and Recommended Mitigations
Maintainers merged a source-code fix on 19 October that forces uploads to resolve inside approved directories and rejects empty names or names beginning with a dot. However, no patched release package is yet available; the newest listed version remains 0.8.5 from 2024.
Until an official release is published, organizations should:
- Block untrusted network access to MONAI Label instances via firewall rules.
- Enable authentication, which is disabled by default.
- Run the server under a least-privilege account rather than root.
- Verify that any source build includes the October path-resolution fix.
- Inspect reachable servers for unexpected Python startup files and anomalous training jobs.
The reported root access occurred on a single university server; outcomes on other deployments depend on runtime configuration and service-account privileges. No evidence indicates that patient scans or clinical systems were accessed.


