Sunday, November 22, 2026
News

Can AI agents really find a zero-day in 11 minutes?

The 11 minutes ended at a file write, and root came in a later session. MONAI Label 0.8.6, released November 22, fixes the flaw, and the old advisory page has not caught up.

Turned on monitor displaying programming language

In one documented case, yes. Amazon says its analysts watched Cairn agents go from mapping a MONAI Label server's API to writing files on it in about 11 minutes. Root came in a later session. MONAI Label 0.8.6, released November 22, fixes the flaw. Versions 0.8.5 and earlier are affected.

What happened in those 11 minutes?

On September 10, 2026, Amazon Threat Intelligence found a public Cairn server with no login. Cairn is an open-source penetration-testing framework driven by AI agents, and the open server showed the operator's instructions and the agents' output. The operator wrote in Simplified Chinese. Amazon places the campaign at about 48 hours from September 8, aimed at university networks in East Asia outside mainland China.

The agents were told to get shell access. Against a medical imaging server they worked out which software it ran, listed its API routes and found an upload route that took a file name from the caller. They tried directory traversal in that name and it worked. Amazon puts about 11 minutes between the API mapping and arbitrary file writes.

The 11 minutes ends at the file write. Root came in a later session, when an agent wrote a Python reverse shell into sitecustomize.py, which Python loads at startup, and then called a training endpoint to launch a fresh Python process. During another session the human operator stepped in and gave the agents a proxy into a second university's network, where Amazon believes the same software was exploited.

Is there a patch, and which MONAI Label versions are affected?

A patch exists as of November 22. MONAI Label 0.8.6 was released that day, and its notes call it a security release that every server deployment should install. It fixes two path traversal flaws that affect every version up to and including 0.8.5. The first is in PUT /datastore/image and PUT /datastore/label, tracked as CVE-2026-133968 and GitHub advisory GHSA-qc22-rmgc-vg94, rated critical. The second is in session upload, PUT /session/?uncompress=true, where tar archives were unpacked without checking member names.

The advisory page still showed no patched version and no CVE when we checked it, so it lags the release. VentureBeat's November 20 report said no patched release existed at that time, and that the fix had been merged on October 19. Authentication is off by default, so any client that can reach the server can use the flaw. With authentication on, a user with the ANNOTATOR role is enough for the datastore bug.

How can I tell whether my MONAI Label server was reached?

Amazon lists two IP addresses seen between September 8 and 10, 142.4.98.145 and 142.4.107.34 (written 142[.]4.98.145 and 142[.]4.107.34 in its post). Search your logs for both.

The release notes tell teams whose 0.8.5 or older server faced an untrusted network to review it for unexpected file changes first. VentureBeat makes the same point, because enabling authentication later does not undo an earlier break-in. Based on the attack description, you can also look for a sitecustomize.py you did not create and for ../ sequences in the image value of PUT requests to /datastore/image.

Do AI agents change how fast teams must patch?

Amazon says they shrink the time defenders have, because an agent that finds and exploits a zero-day in minutes can finish before many security operations centers triage an alert. The same campaign shows limits. Against a Redis server with no authentication, the agents worked for about 16 hours and 148 actions without gaining control. Against a university GPU server they ran for more than 6 hours and tried more than 30,000 credential combinations without finding a working login.

Here the lesson is exposure. A server with authentication off, reachable from an untrusted network, was the weak point. Amazon advises favoring high-confidence detections that trigger an automated response, with people overseeing outcomes instead of reviewing each action.

What should teams running MONAI Label do today?

  • Find every MONAI Label server and check its version. 0.8.5 and earlier are affected.
  • Upgrade to 0.8.6 with pip install -U monailabel==0.8.6 or the 0.8.6 Docker image. It requires Python 3.11 or newer.
  • Turn on authentication by setting MONAI_LABEL_AUTH_ENABLE to True, and give roles only to people you trust.
  • Keep the server off untrusted networks, and run it as an unprivileged user that can write only to its datastore.
  • Search logs and files for the two IP addresses, unexpected file changes, an unfamiliar sitecustomize.py and ../ in image values.

Sources: AWS Security Blog, MONAI Label 0.8.6 release notes, GitHub advisory, VentureBeat, SecurityWeek. Related: Zyxel GS1900: checking for the patch bypass, WorkOS breach: indicators to check.

Sources