Suspected China-linked threat actors utilized autonomous artificial intelligence agents to uncover a previously unknown vulnerability in the open-source medical imaging platform MONAI Label, subsequently leveraging the zero-day to obtain root access to a university server, according to research published by Amazon Threat Intelligence.

The discovery and exploitation took place over approximately 48 hours in September. Notably, the AI tools mapped the application and reached the stage of writing unauthorized files to the target system in around 11 minutes once deployed, as reported by The Hacker News.

Autonomous AI Exploitation via Exposed Cairn Panel

Investigators reconstructed the campaign after discovering the operator's control panel exposed on the public internet without authentication. The dashboard belonged to Cairn, an open-source framework designed to orchestrate AI agents for automated penetration testing. Analysis of the instance exposed configuration settings, task prompts, and operational run logs.

The records revealed that the human operator provided high-level objectives—specifically tasking the tools with obtaining shell access against chosen endpoints—while the agents independently selected and executed the technical discovery methods. The investigators noted with moderate confidence that the activity links to a China-based operator, citing Simplified Chinese system prompts, China Standard Time zone configurations, and targets aligning with strategic interests, though no formal attribution to a state agency or named threat group was established.

Technical Details of CVE-2026-133968

The flaw uncovered by the automated agents, now tracked as CVE-2026-133968 with a critical CVSS score of 9.8, is a path-traversal vulnerability residing in MONAI Label versions 0.8.5 and earlier. MONAI Label is an intelligent image labeling and learning engine used by academic and clinical research teams to annotate medical scans and train specialized AI models.

The vulnerability exists within the application routines responsible for storing uploaded labels and images. When processing incoming requests, the server appends a caller-controlled file identifier directly to the designated storage folder path without validating whether the resolved destination remains confined within the directory boundary.

Because MONAI Label operates without authentication enabled by default, any client capable of reaching the network service can submit crafted requests to write files outside the intended directories. The files are written using the permissions of the underlying service account. While this does not allow arbitrary system-wide writes on its own, planting files into locations referenced by configuration or code execution paths enables arbitrary code execution.

In the observed university intrusion, the root shell was achieved in a subsequent session. The agents exploited the path-traversal weakness to place a malicious file into an environment path that Python automatically imports during startup. The agents then invoked a built-in model training routine, spawning a new Python process that loaded the malicious payload and established a reverse shell with root privileges.

Affected Systems and Current Status

The campaign demonstrated that access levels depend heavily on server configurations. Root access was confirmed on a single university deployment running the server under privileged credentials. Amazon also observed the threat actor providing the agents with network tunnel access into a second university environment utilizing the software, though full compromise of that secondary target remains unconfirmed.

Project maintainers acknowledged the vulnerability and merged a source code patch on October 19. The update enforces strict path resolution within permitted upload directories and explicitly blocks filenames that are empty or begin with a period. However, a packaged, installable release containing the fix has not yet been issued, leaving the latest official release at version 0.8.5.

Amazon's findings indicate no evidence that patient medical records, clinical operational environments, or data beyond the targeted academic research setups were compromised.

What You Should Do

Because a patched distribution package is not yet available, security teams operating MONAI Label deployments should implement immediate mitigation measures:

  • Restrict Network Reachability: Place MONAI Label instances behind network firewalls, VPNs, or private subnets to block untrusted external access entirely rather than relying on obfuscated URLs.
  • Enable Authentication: Activate application authentication controls, as the software ships with access verification disabled by default.
  • Enforce Least Privilege: Transition MONAI Label services away from privileged or root accounts to a dedicated service user with strictly restricted write permissions on the local filesystem.
  • Audit System Directories: Inspect running environments for anomalous Python initialization scripts, unexpected startup files, or unscheduled model training tasks.
  • Verify Source Builds: If building the application directly from source code, ensure the commit history includes the directory resolution fixes merged on October 19 before deploying into production.