Skip to content

ResourcesNews

Amazon says hackers' AI agents found a zero-day on their own

Agents run by a suspected China-linked group went from mapping a medical imaging app to writing files on its server in about 11 minutes.

By Jason NguyenPublished 2 min read
A small white robot shines a flashlight on a keyhole in a plain box on a dusty blue background
Image: Illustration / Recaply

In short

  1. 01Amazon Threat Intelligence spotted the campaign after the operator left a Cairn server open to the internet with no login.
  2. 02MONAI Label 0.8.6 fixes the critical flaw, tracked as CVE-2026-133968, and every earlier version is affected.
  3. 03Amazon links the operator to China with moderate confidence, based on Simplified Chinese prompts, China Standard Time timestamps and the choice of targets.

The story

Amazon just published a threat report on hackers who gave AI agents a target list and let them work. The agents found an unknown flaw in MONAI Label, an open-source medical imaging tool, and took full control of a server.

The operator's goal for the agents was short, to "obtain shell access." They ran on Cairn, an open-source penetration-testing framework built for defenders. BleepingComputer reports that Cairn has also appeared in a credit card theft operation and an attack on a U.S. aerospace organization.

Two long attacks failed, but another found the flaw

The campaign hit universities in East Asia outside mainland China over about 48 hours. At one school, the agents spent roughly 16 hours and 148 actions trying to take over a Redis database server. They never got in.

A second run lasted more than six hours. The agents built a list of about 80 real usernames from public university pages, then tested more than 30,000 login combinations. None worked.

The medical imaging target went differently. In about 11 minutes, the agents moved from mapping the app's API to writing files on its server. A later session turned that into root access, the highest level of control.

The operator then stepped in and gave the agents a tunnel into a second university network. Amazon infers the same flaw was used to break into the server at the other end.

The 11-minute figure comes from Amazon's own review of the exposed records. Recaply couldn't find an independent analysis of them.

MONAI Label's fix took weeks to reach a release

Maintainers merged the code fix about five weeks before they shipped it. Version 0.8.6, the first release to include it, landed after Amazon's report. The previous release was two years old.

The project's security advisory rates the bug critical and warns that login is off by default. That means anyone who can reach an unpatched server over the network can exploit it.

Version 0.8.6 is out now on pip and Docker. Teams that can't upgrade yet should turn on authentication and keep the server off untrusted networks.

Why it matters

Agents make patient attacks possible against systems nobody watches closely. This flaw sat in a niche research tool, and the agents found it without being told it existed. Universities and labs should start by finding research servers with login turned off, since an agent can keep probing one for 16 hours.

Get Recaply Security in your inboxFree, unsubscribe any time.

More stories