Skip to content
OriginalAnalysisSecurity

Attackers' AI agents found a zero-day in 11 minutes

Autonomous AI agents run by a suspected China-linked operator found an unknown flaw in MONAI Label, a medical imaging server, and used it to gain root on a university server. No fixed release exists yet.

A chalk maze drawn on a dark slate blackboard, with a single ochre-yellow chalk line threading through the maze and escaping out the bottom-right edge
A chalk maze on a slate board, one ochre line finding the way out: the path the agents found through the file system.
On this page

In brief

  • Autonomous AI agents run by a suspected China-linked operator found an unknown flaw in MONAI Label and used it to gain root on a university server.
  • The flaw, CVE-2026-133968, is a path traversal bug rated 9.8. It lets anyone who can reach the server write files outside its storage folder.
  • Amazon found the campaign because the operator left the agents' own Cairn control panel open to the internet with no password.
  • A fix was merged on Oct 19, but no patched release exists yet. Version 0.8.5 from 2024 is still the newest package.
  • Until a release ships, the defense is to keep MONAI Label off untrusted networks, turn on authentication, and run it as an unprivileged user.

The people running the attack left their own dashboard unlocked. That is how Amazon found them.

On September 10, Amazon Threat Intelligence came across a server on the public internet running Cairn, an open-source framework that runs AI agents as automated penetration testers, with no password on it. The control panel was open to anyone: the operator's project settings, the prompts the agents were given, and the full history of what they had done. Amazon could read the campaign the way its operator could.

What it showed was a two-day operation against universities in East Asia, beginning September 8, in which autonomous agents did the hacking with almost no human direction. The operator gave them a list of target addresses and a goal in plain language, "obtain shell access," and the agents planned each next step themselves. In one case, they found a flaw nobody knew about and used it to take full control of a server. The flaw is now tracked as CVE-2026-133968, rated 9.8 out of 10.

The target was MONAI Label, an open-source server that research teams use to label medical scans and train imaging models. A lab installs it, points it at a folder of scans, and annotators mark up images through a web interface or an API. It is a niche tool, built for research groups, and it ships with authentication switched off. Anyone who can reach the server on the network can use it.

What the agents found

The agents mapped the application's programming interface, found an upload endpoint that takes a file name from the caller, and tested whether that name could escape the folder it was supposed to stay in. It could. MONAI Label took the name it was given, joined it to its storage folder, and wrote the uploaded file to the resulting path without checking that the path stayed inside the folder. A name like ../../../tmp/x would land the file wherever the server process had permission to write. In security terms this is a path traversal flaw. In plain terms, the server trusted the label on the envelope and filed the letter wherever the label said.

Eleven minutes passed between the agents mapping the interface and writing files anywhere on the server, according to Amazon. Eleven minutes is the time from first mapping the API to arbitrary file writes, not the time to full control. The root shell came in a later session, when the agents used the same flaw to plant a small program in sitecustomize.py, a file Python loads automatically every time it starts, then triggered the application's training feature to start a new Python process. The new process loaded the planted file and connected back to the attackers with root privileges.

The failures matter too

The agents were persistent in a way a human operator is not, and they failed most of the time. One operation ran for about 16 hours and 148 sequential actions against a Redis database left open without authentication at one university, working through replication abuse, sandbox escapes, configuration writes, and memory corruption, then mapping the internal Docker network through a flaw in a neighboring service and spraying credentials at PostgreSQL, SSH, and other services. None of it worked. A second operation spent more than six hours on a university GPU server, building a list of about 80 real usernames from faculty directories, lab sites, archived pages, and papers, then trying more than 30,000 password combinations against its remote login and other systems across the campus. It never found a working login.

That is the honest shape of the story: agents that keep going for a day without sleep or working hours, fail repeatedly, and then find the one hole that matters. The zero-day was not the product of a brilliant plan. It was the product of trying everything, quickly, until something gave.

Amazon attributes the campaign to a Chinese-nexus actor with moderate confidence, based on prompts written in Simplified Chinese, scripts timestamped to China Standard Time, and targets likely to interest China. That is not a named group, and Amazon does not claim a government was behind it. The assessment is a judgment about language, time zones, and target selection, and it should be read as one.

A fix with no release

The flaw itself is now public, and the fix is in an awkward state. Project MONAI's maintainers confirmed the flaw and merged a fix in pull request 1915 on October 19. Their security advisory, GHSA-qc22-rmgc-vg94, says the correction will be included in the next release. There is no patched release to install yet. The newest package remains 0.8.5, which dates to 2024. A fix in the source code is not the same as a package you can install, and the advisory is blunt about what to do in the meantime: switch authentication on, keep the server off untrusted networks, and run it as an unprivileged user whose write access stops at the storage folder. Even with authentication on, the advisory notes, a user with the annotator role can reach the vulnerable endpoints.

Two wider points follow. The first is that the framework the attackers used was built for defenders. Cairn is open source, designed as a research engine for finding weaknesses before attackers do. Amazon notes that a financially motivated group was already using it against online retailers, documented earlier by Gambit Security. The same tool, pointed at different targets, serves both sides, and it costs its users almost nothing to run at machine speed.

The second is what the disclosure says about response time. An agent that finds and exploits a flaw in minutes finishes before most security teams have triaged the alert. Amazon's own advice is to lean on high-confidence detections that trigger automated responses, with people overseeing outcomes rather than reviewing every action. That is a frank admission: the checking still runs at human speed, and the finding no longer does.

If you run MONAI Label, the steps are unglamorous and they are the defense until a release ships. Block untrusted network access to the server. Turn on authentication. Run it as a limited account, not root. And if the server was ever reachable, check for unexpected Python startup files and unusual training jobs, the exact path the reported attack used. There is no patch to install yet, so the network boundary is the patch for now.

Related reporting

More from grep.tips

All articles