Quick Facts

  • A Chinese-nexus threat actor used autonomous AI agents running the open-source Cairn framework to find and exploit a zero-day vulnerability in MONAI Label, a medical imaging application, in approximately 11 minutes.
  • The flaw, CVE-2026-133968, carries a CVSS score of 9.8 and allows arbitrary file writes that can lead to root access. A fix was merged October 19, but no patched release has shipped.
  • Amazon Threat Intelligence discovered the campaign on September 10, 2026, after the operator left a Cairn instance publicly exposed with no authentication, making its full operational history visible.

Amazon Threat Intelligence disclosed on November 20, 2026, that a Chinese-nexus threat actor had used autonomous AI agents to find and exploit a previously unknown vulnerability in an open-source server application, gaining full root access with minimal human direction.

The attack targeted a university server in East Asia running MONAI Label, an open-source medical imaging AI tool. The agents mapped the application's API endpoints, identified a vulnerable upload endpoint, and wrote arbitrary files to the server in about 11 minutes.

How the Attack Worked

The threat actor used Cairn, a legitimate open-source penetration testing framework, to coordinate multiple AI agents through a shared blackboard system. Every confirmed discovery became a fixed fact; every planned move became a queued action. The operator gave the agents a single goal: obtain shell access.

The agents located an upload endpoint in MONAI Label where the filename parameter was vulnerable to path traversal. The underlying flaw, now tracked as CVE-2026-133968 and published in GitHub Security Advisory GHSA-qc22-rmgc-vg94, exists because the application's LocalDatastore.add_image() function never verifies that a resolved file path stays within the datastore directory. A value such as ../../../tmp/x writes content anywhere the MONAI Label process has write access.

In a follow-up session, the agents wrote a Python reverse shell to sitecustomize.py, a file Python loads automatically on startup, then triggered an API endpoint to establish root access. By default, MONAI Label runs with authentication disabled, leaving any network-reachable instance immediately exposed.

The Operator Left the Door Open

Amazon's researchers found the campaign only because the threat actor made a significant operational mistake. They left their Cairn instance publicly accessible with no authentication. That exposed the platform's full history, including the operator's prompts, agent outputs, and project configurations.

Analysis of that history showed the agents ran continuously for 16 hours without breaks or human direction across multiple targets, including two unsuccessful attempts against a Redis instance and a university GPU server. The operator later gave the agents a proxy into a second university network. Amazon assesses the operator had already used the same zero-day to compromise that server and build the tunnel.

Cairn's author, who posts as "le1xia0," told Reuters the tool was not built for attacks and that harmful use was not the intent.

No Patch Is Available Yet

The MONAI Label project published the GitHub Security Advisory on November 18, 2026, rating the vulnerability Critical. A code fix was merged on October 19, but version 0.8.5 remains the newest release. The patch will ship in the next MONAI Label release. Until then, the advisory's recommended workaround is to block all untrusted network access to any MONAI Label server.

Both add_image() and save_label() will resolve the destination path and reject any location outside the designated datastore directories once the patched version ships.

What This Means for Security Teams

Amazon's Chief Information Security Officer CJ Moses wrote that an agent completing exploitation faster than a security operations center can triage an alert represents a new speed problem for defenders. He called on organizations to prioritize high-confidence detections that trigger automatic responses, with humans monitoring outcomes rather than approving each individual action.

Moses also noted that the successful attacks shared common conditions: default or hardcoded credentials, keys left in public artifacts, unauthenticated interfaces, and unpatched software. He wrote that getting those fundamentals right remains the most effective defense against AI-assisted attacks.

Amazon assessed with high confidence that the MONAI Label flaw was a zero-day at the time of exploitation, citing no corresponding fix, public issue, advisory, or vulnerability database entry at that point. The campaign ran across a 48-hour window starting September 8, 2026.

This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.