A futuristic medical lab depicts AI-assisted imaging data moving to servers, with cybersecurity icons highlighting privacy and network risks.
Autonomous hacking agents have just posted a result that defenders can't wave off. Amazon Threat Intelligence says an AI agent framework called Cairn found a previously unknown flaw in a medical-imaging application and used it to take over a server. The operator's instruction was reportedly just "obtain shell access." Amazon says Cairn's agents found a path traversal zero-day in MONAI Label, an open-source medical imaging application that uses AI to help researchers label images.

Windows admins shouldn't dismiss this as a Linux-and-Python curiosity. The same pattern can hit any internet-reachable service with weak defaults. The details below cover what happened, what failed, which versions are affected, and what to do now.

A futuristic medical lab depicts AI-assisted imaging data moving to servers, with cybersecurity icons highlighting privacy and network risks. What Amazon says happened​

On September 10, 2026, Amazon found a publicly reachable Cairn server running without authentication. Cairn is an open-source autonomous penetration-testing framework. Because the operator left it open, anyone could see its configuration, prompts and agent output. Amazon says the prompts were in Simplified Chinese and the script timestamps used China Standard Time.

The activity ran for roughly 48 hours starting September 8. It targeted IP addresses associated with universities in East Asia outside mainland China. Amazon's CISO-level report describes the campaign as aimed at universities in East Asia outside Mainland China. Amazon did not name the universities or the country. It says the organizations weren't AWS customers, so it worked with a government partner to notify them.

The operator supplied target IPs and plain-language goals, and the agents planned and adjusted their approach on their own. Amazon says they did this "without the operator choosing each next step."

The attack that worked​

Against the medical imaging application, the agents followed a short, methodical chain:

  1. They identified the application framework and mapped its API endpoints.
  2. They found an upload endpoint with a filename parameter and tested it for path traversal.
  3. About 11 minutes after starting API mapping, they could write files to arbitrary locations on the server.
  4. In a later session, they used the flaw to write a Python reverse-shell payload into sitecustomize.py. Python imports that file automatically at interpreter startup.
  5. They triggered a training API endpoint, which spawned a new Python process. That process loaded the payload and opened a reverse shell running as root.

Overwriting sitecustomize.py is a neat trick because it needs no exploit chain. It only needs a file-write primitive and a way to make Python start. Any Python service that restarts workers or spawns jobs gives an attacker that second condition.

Amazon says the operator later gave the agents a proxy into a second university network. The server behind that proxy was running the same vulnerable software, and Amazon infers the operator had used the zero-day to compromise it and build the tunnel. The operator then pointed the agents at other machines in that network. The public account doesn't say those additional attempts succeeded.

The failures matter just as much​

This was not a flawless machine. The agents failed at several other tasks:

  • Redis target: About 16 hours and 148 sequential actions went into an unauthenticated Redis instance. BleepingComputer notes Cairn was built as a legitimate tool for research and defense. The agents tried Redis replication abuse, scripting-sandbox escape, config-file writes, memory corruption and Log4Shell against another service on the same host. They then mapped an internal Docker network through an SSRF flaw in a co-located Qdrant instance and sprayed credentials at PostgreSQL, SSH and other services. Nothing worked.
  • GPU server: For more than six hours, the agents built a list of about 80 real usernames from faculty directories, lab sites, archived pages and academic papers. They then tried over 30,000 credential combinations against SSH, an admin dashboard, network storage, a VPN and a campus portal. They found no working login.

The takeaway is that hardened targets held, but the agents never tired and never stopped adapting. Amazon points out that the Redis effort ran continuously "without breaks, sleep, or working-hour constraints." Meanwhile, the one soft target, a default-configured app with an unpatched bug, fell in minutes.

The vulnerability: affected versions and status​

The MONAI Label project published GitHub Security Advisory GHSA-qc22-rmgc-vg94 on November 18, 2026, rated Critical. Key facts from the advisory:

ItemDetail
Packagemonailabel (pip)
Affected versions0.8.5 and earlier
Patched versionsListed as "None"
CVENo known CVE ID
WeaknessCWE-22, path traversal
FixPull request #1915, to be included in the next release

There is a discrepancy on patch status. BleepingComputer says maintainers released a security update to address the flaw on November 18. The advisory itself lists no patched package version and says the fix will ship in the next MONAI Label release. Until a release lands, treat the code fix as merged but not yet packaged. Check the project's release page before assuming pip install --upgrade solves it.

How the bug works​

The PUT /datastore/image endpoint takes an image query parameter and uses it as the image ID without validation. The application adds a file extension, joins the value onto the datastore image directory and writes the uploaded file. It never checks that the resulting path stays inside the datastore. A value like ../../../tmp/x therefore writes uploaded content anywhere the MONAI Label process can write. The PUT /datastore/label endpoint builds its destination from the image ID and label tag in a similar way.

The advisory says arbitrary file write can lead to remote code execution, for example by overwriting application code or adding entries to .ssh/authorized_keys or cron. It can also cause denial of service by overwriting configuration files.

The default configuration makes this worse. Authentication is off by default (MONAI_LABEL_AUTH_ENABLE=False), so any network client that can reach a default server can exploit it. With authentication enabled, the ANNOTATOR role is enough.

What to do now​

If you run or support MONAI Label, including research groups who stood one up on a GPU box, the maintainers' workarounds are:

  1. Set MONAI_LABEL_AUTH_ENABLE=True and don't grant the ANNOTATOR role to untrusted users.
  2. Don't expose the server to untrusted networks.
  3. Run it as an unprivileged user whose write access is limited to the datastore directory.
  4. Watch for the next release that includes pull request #1915 and upgrade when it appears.

Hunting steps follow from the attack chain, though they are general practice rather than Amazon guidance:

  • Look for unexpected or recently modified sitecustomize.py files in Python site-packages directories.
  • Check for unexpected entries in authorized_keys and cron files.
  • Review web logs for PUT requests to the datastore endpoints with .. sequences in the image parameter.
  • Check outbound connections from the server for unexplained reverse-shell traffic.
  • Search firewall and proxy logs for the two published indicators, 142[.]4.98.145 and 142[.]4.107.34. Amazon says these appear in records dated September 8 to 10, 2026. They are reported indicators, so treat a hit as a lead to investigate rather than proof.

Attribution: how much to read into it​

Amazon assesses with moderate confidence that the actor is Chinese-nexus. It bases that on the Simplified Chinese prompts, the China Standard Time timestamps and the targeting of universities likely of strategic interest to China. Those are circumstantial clues. They don't identify a specific group or a government sponsor, and Amazon doesn't claim they do. By contrast, Amazon assesses with high confidence that the flaw was a zero-day. It reached that view because it found no fix, issue, advisory or prior mention in vulnerability databases. Neither the advisory nor Amazon's account claims data theft.

The bigger picture​

Cairn is no longer a one-off. Per BleepingComputer, Gambit Security documented its abuse in a large-scale credit card theft operation in September, and Google reported it in an attack on a US aerospace organization. Gambit's Eyal Sela told The Record that Cairn is a general-purpose tool with no internal cybersecurity structuring, which makes it less predictable. Amazon's own warning is blunt. An agent that exploits a zero-day in minutes finishes before many security operations centers can triage an alert.

Amazon's recommended response is to prioritize high-confidence detections that trigger automated response, with humans overseeing outcomes rather than reviewing every action. That advice has limits. Automated response needs good detections, and a bad auto-block rule can take down your own service. Still, the arithmetic is hard to argue with. Defenders can't hand-review every alert at machine speed.

For Windows and enterprise admins, three practical lessons stand out:

  • Defaults are the attack surface. An authentication-off default plus a network listener is exactly what an agent goes looking for.
  • Research and lab servers are real targets. GPU boxes and departmental apps often sit outside normal patch management.
  • Least privilege limits the damage. A process that can only write to its datastore can't plant a root-level reverse shell.

The agents lost against the hardened Redis host and the GPU server's logins. They won against an unpatched app with no authentication. Basic hygiene still works, but only if you apply it to the servers nobody remembers they deployed.

 

References

  1. Hackers use autonomous AI tool Cairn to find zero-day, take over server BleepingComputer 2026-11-20T12:30:00-05:00
  2. Path traversal in datastore upload endpoints allows arbitrary file write · Advisory · Project-MONAI/MONAILabel · GitHub github.com
  3. Chinese-nexus actor’s Cairn AI agents found a zero-day on their own | AWS Security Blog aws.amazon.com