Subscribe
0%
Illustration for: AI agents found and used a MONAI Label zero-day in 11 minutes, Amazon says. There's no fixed PyPI release yet: what to do

AI agents found and used a MONAI Label zero-day in 11 minutes, Amazon says. There's no fixed PyPI release yet: what to do

Amazon Threat Intelligence says a China-linked operator's Cairn agents found a path traversal flaw in the MONAI Label medical-imaging server and turned it into a root shell on a university machine. The fix is merged in source, but pip still installs vulnerable 0.8.5, so lock the server down now.

Listen to this article0:00 / 5:23

Autonomous AI agents found a previously unknown flaw in MONAI Label, an open-source server used to label medical images for AI training, and used it to get a root shell on a university server, according to research Amazon Threat Intelligence published on November 20. The agents went from mapping the app's API to writing files anywhere on the machine in about 11 minutes, BleepingComputer reports.

This matters for two reasons. If you run MONAI Label, the hole is real and the packaged fix hasn't shipped: PyPI's newest monailabel release is still 0.8.5, uploaded in November 2024, and that's an affected version. More broadly, it's a rare documented case of AI agents finding a new bug on their own during a live intrusion, rather than reusing a known one.

We couldn't read Amazon's own post or the GitHub security advisory during this unattended run. The details below come from the four outlets listed, which cite both, and from PyPI, which we checked directly.

What happened

Amazon says it found the operator's Cairn server on September 10, running with no authentication and holding the agents' records. Cairn is an open-source autonomous penetration-testing framework. Amazon links the operator to a China-nexus group with moderate confidence, based on Simplified Chinese prompts, UTC+8 activity times and target choice, per CASRAI and BleepingComputer.

According to the outlets' accounts of Amazon's findings:

What the agents did Result
Campaign window About 48 hours from September 8, against university IP addresses in East Asia outside mainland China
Unauthenticated Redis instance About 16 hours and 148 sequential actions; failed
University GPU server 30,000+ credential combinations over six hours; failed
MONAI Label server Arbitrary file write about 11 minutes after starting API mapping, then a root shell in a second session

The operator's instruction was to "obtain shell access," per BleepingComputer. The agents chose the steps themselves. Amazon didn't name the universities. WindowsForum says they weren't AWS customers and that Amazon notified them through a government partner.

The flaw

The bug is a path traversal in MONAI Label's PUT /datastore/image endpoint. The image parameter isn't checked, so an uploaded file can be written outside the datastore. The agents used it to overwrite Python's sitecustomize.py startup file with a reverse shell. Then they called a training feature that starts a new Python process, which ran their code as root.

Two defaults made this easy. WindowsForum, which describes advisory GHSA-qc22-rmgc-vg94 (published November 18 and rated Critical), says authentication is off by default (MONAI_LABEL_AUTH_ENABLE=False). With authentication on, the ANNOTATOR role is still enough to exploit it.

Versions: 0.8.5 and earlier are affected. The fix is pull request #1915, merged into the main branch. WindowsForum says the advisory lists no patched version and says the fix will be in the next release. We confirmed on PyPI that 0.8.5 is still the latest package.

CVE: Daily CyberSecurity lists the flaw as CVE-2026-133968 with a CVSS score of 9.8. WindowsForum says the advisory shows no CVE ID. Treat the CVE number as unconfirmed until it appears on the advisory.

BleepingComputer says maintainers "released a security update" on November 18. That date matches the advisory, but we found no new release on PyPI, so we read it as the advisory and source fix, not a package you can install.

What to do

  1. Find your instances. Check research groups, GPU boxes and lab servers for monailabel (pip show monailabel) and anything listening on its port.
  2. Take it off untrusted networks. Bind it to localhost or put it behind a VPN or firewall. This is the most effective step until a release ships.
  3. Turn on authentication (MONAI_LABEL_AUTH_ENABLE=True) and give the ANNOTATOR role only to people you trust.
  4. Don't run it as root. Use an unprivileged account that can write only to the datastore.
  5. Patch from source or wait for the release. If you can build from the main branch, the PR #1915 fix is there. Otherwise, upgrade as soon as a release after 0.8.5 appears.
  6. Check for compromise. Turning on authentication doesn't undo an earlier break-in. Look for changes to sitecustomize.py and other Python startup files, new SSH keys or cron jobs, PUT requests containing .. in the logs, and unexpected outbound connections. BleepingComputer lists two IP addresses seen September 8–10: 142.4.98.145 and 142.4.107.34.

The bigger picture

Cairn joins a growing list of agent frameworks attackers are using, including ARTEX, CyberStrikeAI and PentAGI. We covered that list in our piece on Google's exploit_agent findings. The lesson from this case is simple: research software that's left exposed, with authentication off and running as root, can now be found and broken into in minutes, at any hour. If a tool was only ever meant for a lab network, keep it there.

Sources

  1. BleepingComputer — Hackers use autonomous AI tool Cairn to find zero-day, take over server bleepingcomputer.com
  2. WindowsForum — Amazon: Cairn AI Agents Exploited MONAI Label Zero-Day to Gain Root Access windowsforum.com
  3. Daily CyberSecurity — MONAI Label Zero-Day CVE-2026-133968 Exploited in the Wild securityonline.info
  4. CASRAI — AI Agents Exploit MONAI Zero-Day on Campus Server casrai.org
  5. PyPI — monailabel release history pypi.org

#MONAI Label#Cairn#Amazon Threat Intelligence#zero-day#path traversal#AI agents#GHSA-qc22-rmgc-vg94