INC-2026-0009SEV3 · MediumAgent actionConfirmed

An attacker’s Cairn AI agents find a zero-day in a medical imaging tool and gain root

A suspected Chinese-nexus actor gave Cairn, an open-source AI penetration-testing framework, target addresses at East Asian universities and the goal to obtain shell access. Amazon reports that the agents found an unknown flaw in a medical imaging AI application and gained root on a server.

Occurred
Disclosed
Published
System
Cairn autonomous AI penetration-testing framework, run by a suspected Chinese-nexus threat actor
Impact
Root access on a server through a flaw the agents found themselves. Amazon infers that a server in a second university network was compromised the same way.
Root cause
An unpatched path traversal flaw in the application’s upload endpoint let the agents write files to any location on the server.

Timeline

  • 2026-09-08. A campaign against IP addresses associated with universities in East Asia outside mainland China begins and runs for about 48 hours, according to Amazon.
  • 2026-09-10. Amazon Threat Intelligence identifies a server hosting a Cairn instance without authentication. Its project configurations, the operator’s prompts and the agents’ output are visible to anyone.
  • 2026-11-18. The MONAI Label maintainers publish an advisory for a path traversal flaw in the datastore upload endpoints.
  • 2026-11-20. Amazon publishes its findings. BleepingComputer reports them the same day.

Trigger or entry point

According to Amazon, the operator gave Cairn a list of target IP addresses and variations of the goal “obtain shell access”. The agents planned, carried out and adjusted their approach without the operator choosing each next step. Against a server running an open-source medical imaging AI application, they identified its framework, mapped its API and found an upload endpoint with a filename parameter that was open to path traversal. Amazon does not name the application. BleepingComputer identifies it as MONAI Label and links to the maintainers’ advisory.

Failure mechanism

Amazon reports that the agents went from mapping the API to writing arbitrary files on the server in about 11 minutes. In a later session they used the flaw to place code that Python loads at startup, then called a training endpoint to start a new Python process, which gave them a shell running as root. Amazon assesses with high confidence that the flaw was a zero-day, because no fix, issue or advisory for it existed. The maintainers’ advisory says authentication is off by default, so in the default configuration any client that can reach the server can exploit the flaw. Amazon does not say how the targeted server was configured.

Impact

One server was taken over with root access. In another session the operator gave the agents a proxy into a second university network, whose server ran the same software. Amazon infers that the operator had compromised that server with the same flaw. Two other operations failed: about 16 hours and 148 actions against an unauthenticated Redis server at one university, and more than 30,000 credential combinations over more than 6 hours against a university GPU server and other systems. Amazon’s report does not mention stolen data. Amazon attributes the campaign with moderate confidence to a Chinese-nexus actor, based on the prompt language, the script timestamps and the choice of targets.

Response

Amazon notified the maintainers, who fixed the flaw and published a GitHub Security Advisory. The advisory says the fix will be part of the next MONAI Label release. The affected organizations were not AWS customers, so AWS worked with a government partner to notify them.

Mitigations

Until an upgrade is available, the advisory recommends turning on authentication, not giving the annotator role to untrusted users and keeping the server off untrusted networks. Amazon recommends high-confidence detections that trigger an automated response, with people overseeing the outcomes, because agents act faster than analysts can triage alerts.

Engineering takeaway

Failure pattern
Attack agents given only targets and a goal found an unknown path traversal flaw in an exposed research application and turned it into root access, with no human choosing the steps.
Preventive control
Keep research and AI tools such as labeling and training servers off the internet, turn on their authentication, run them without root, and alert when a service writes to Python startup files.
Transferable lesson
Agent attackers go from first request to compromise in minutes and keep trying for hours, so exposed services need automated detection and response rather than manual triage.

Editorial summary by Incidents.si, drawn from the sourced record above. It is not a statement by the organizations involved.

Sources

  1. Chinese-nexus actor’s Cairn AI agents found a zero-day on their ownprimary · 2026-11-20
  2. Path traversal in datastore upload endpoints allows arbitrary file writeprimary · 2026-11-18
  3. Hackers use autonomous AI tool Cairn to find zero-day, take over serversecondary · 2026-11-20

Corrections

No corrections have been made to this entry.

Report a correction