Amazon Threat Intelligence says China-linked attackers used autonomous AI agents to find and exploit a MONAI Label zero-day on a university server, then obtained root. The agents ran on Cairn, an open-source AI penetration-testing framework, according to SecurityWeek. Amazon published its findings on November 20, 2026, and coverage continued through November 21.

The case stands out for two reasons. Amazon's report describes agents discovering a new flaw during a live intrusion, not just chaining bugs that were already known. A Cloud Security Alliance briefing describes this as a move beyond earlier AI-assisted attacks. The software also sits in medical-imaging research, where US hospitals, labs and vendors may run it.

How the MONAI Label zero-day was exploited

MONAI Label is an open-source server for AI-assisted annotation of medical images, according to the Cloud Security Alliance research note. The flaw is a path-traversal bug in the application's file-upload function. Path traversal lets a request escape the directory an application intends to use, so it can write files in places it should not reach.

The agents mapped the application's API first. About 11 minutes later, they were writing files anywhere on the server, according to TechBytes.

The second stage was more consequential. In a separate session, the agents placed a Python reverse shell in sitecustomize.py, a file Python loads at startup, and obtained root, according to Cryptopolitan. A reverse shell makes the compromised machine connect back to the attacker, which helps an intruder get around inbound firewall rules. Amazon also said the agents compromised a second machine that the attackers used to build a tunnel.

The operator reportedly gave the agents one goal: 'obtain shell access,' per WindowsForum. The instruction is short, which fits the autonomous approach Amazon describes. Amazon said the campaign came to its attention after it found an exposed Cairn instance running without authentication, per PRSOL:CC.

Key dates and affected MONAI Label versions

The dates in the reporting do not all match. The table lists the ones that appear across the sources.

Date What the sources report
September 8 to 10, 2026 Activity window, based on IP addresses Amazon shared
September 2026 Amazon detects the activity
October 19, 2026 Fix-merge date, per one report
November 18, 2026 Maintainer security update, per another report
November 20, 2026 Amazon publishes its findings

Sources identify MONAI Label 0.8.5 and earlier as affected, according to WindowsForum and TechBytes. Any installation at or below that version should be treated as exposed until its release status is confirmed. Research servers set up for one project and then left alone are easy to miss, so start the inventory there.

Patch status is disputed, so verify before upgrading

Patch status is the most confusing part of this story, and the sources disagree. One report says the advisory lists no patched package version and that the fix will ship in the next release, per WindowsForum. Another says maintainers released a security update on November 18, per PRSOL:CC. The Cloud Security Alliance briefing says no patched release exists yet, while SecurityOnline tells administrators to install the fix right away. Treat none of those claims as final until the official MONAI advisory names the patched version.

SecurityOnline is the only source in the brief that names a CVE, CVE-2026-133968, and it gives a CVSS score of 9.8. The Cloud Security Alliance briefing cites GitHub advisory GHSA-qc22-rmgc-vg94. Confirm both identifiers against the official advisory, because the CVE has not been independently confirmed in the coverage reviewed for this story. Amazon's vulnerability team reported the flaw to Project MONAI, according to the CSA research note.

What to do if you run MONAI Label

The immediate steps are clear, even though the patch details are not settled.

  1. Find every MONAI Label installation, including research and test servers, and record the exact version.
  2. Flag anything at version 0.8.5 or earlier for immediate review.
  3. Restrict network access so MONAI Label cannot be reached from the internet, and turn on authentication.
  4. Install the Project MONAI fix once the official advisory confirms the patched release.
  5. Look for exposed, unauthenticated AI agent tools in your own environment, including Cairn instances, because Amazon's own discovery began with one.
  6. Let high-confidence detections trigger automated responses, with people reviewing the outcomes. Amazon urged this approach, according to Cryptopolitan.

What Amazon has and has not proven

Amazon links the operation to a China-linked or Chinese-nexus actor. That is a suspicion rather than a formal attribution, and the sources reviewed do not confirm it independently. Amazon also assessed 'with high confidence' that the bug was a zero-day, since no fix, advisory or prior mention existed, per SecurityWeek. Both conclusions rest on Amazon's analysis.

The reporting also leaves the victim unnamed. Sources describe a single university server in East Asia, and none identifies a US victim. Any US exposure is an inference from where MONAI Label is used, not a reported incident.

Why AI agents change the defense math

Cairn is available to defenders as well as attackers, and that is part of why Amazon's report matters. The same class of tooling that helped an operator run this intrusion can also help security teams test their own systems before someone else does. SecurityOnline quoted the report's bottom line: 'These findings show that agentic offensive operations are already happening.'

The time window is what makes this hard. An 11-minute gap between mapping an API and writing arbitrary files leaves little room for a ticket queue, a manual review or a weekend change freeze. That is why Amazon's advice to automate responses to high-confidence detections carries weight. It also means security teams need to trust those detections before they hand software control over containment.

What to watch next

Several open questions will determine how serious this is for US healthcare and research teams. Watch for an official Project MONAI advisory that names a patched version and resolves the conflicting October 19 and November 18 dates. Watch for a second source that confirms CVE-2026-133968 and its CVSS score. Watch also for other reports of AI agents finding zero-days during live intrusions, which would show whether this case is a pattern or an early outlier. Amazon's own write-up would answer many of these questions, and readers should check it directly when they can.