The MONAI Label zero-day became the decisive step in a Cairn campaign against East Asian universities. Amazon Threat Intelligence says autonomous agents found an arbitrary file write in the medical imaging application, then used it to gain root access on one server. For teams evaluating an external attack surface management platform, the case raises a concrete inventory question: which research services are reachable from untrusted networks?
Amazon published its investigation on November 20, 2026, after observing activity that began in September. The MONAI Label maintainer advisory, published November 18, identifies the affected package as monailabel 0.8.5 and earlier. The maintainers have fixed the code, but the advisory still lists no patched release.
How the MONAI Label Zero-Day Changed the Campaign
Amazon says it found an unauthenticated Cairn server on September 10. Its exposed operational history showed prompts, projects, and agent output from an approximately 48-hour campaign that began September 8. The targets were university IP addresses in East Asia outside mainland China, according to Amazon.
The operator supplied target addresses and a broad objective to obtain shell access. Cairn coordinated several agents that chose and revised their own next actions. Amazon observed both lengthy unsuccessful efforts and one successful compromise. These outcomes must be kept separate when assessing the campaign.
At one university, the agents spent about 16 hours and 148 sequential actions against an accessible Redis service and nearby systems. They tried multiple routes and mapped other reachable services, yet Amazon says none succeeded. A separate six-hour effort tested more than 30,000 credential combinations against university systems after assembling about 80 real usernames; it also failed.
Those counts describe attempted activity in Amazon’s observation, not successful access or a broader victim total. The successful MONAI Label operation followed a different path. Amazon says agents moved from identifying the application’s API to an arbitrary file write in about 11 minutes. In a later session, they used the weakness to obtain a root shell.
- Reachable serviceInternetapplication identified
- MONAI LabelApplicationfile write weakness
- Application processHostnew process started
- Server controlHost
This diagram summarizes Amazon’s observed sequence without reproducing the operational method. A reachable service gave the agents a place to test the application. The write weakness crossed the boundary between application data and server files. The resulting host access was the confirmed outcome in this operation.
Why Could a File Upload Become Server Control?
The maintainer advisory describes path traversal in MONAI Label datastore uploads. A caller-controlled identifier could influence where uploaded content was written, and the application did not ensure the destination stayed inside the datastore. The consequence was an arbitrary write wherever the MONAI Label process had filesystem permission.
That distinction explains why exposure and process privileges matter together. An arbitrary write can damage application files or place content where another process later uses it. Amazon reports that the agents used the flaw in a later session and caused a newly started Python process to execute their content. The resulting shell ran as root in Amazon’s observed case.
The advisory says authentication is disabled by default. In that configuration, any network client able to reach the server can attempt the flaw. If authentication is enabled, a user with the ANNOTATOR role has sufficient access. A login screen alone is therefore not the whole exposure test; role grants and network reachability also matter.
MONAI Label 0.8.5 and earlier are listed as affected. The advisory marks patched versions as “None,” while stating that a fix landed in the project and will enter its next release. Administrators should distinguish a merged source change from a packaged version they have actually installed. The source does not give a released fixed version to recommend today.
What Does the Campaign Establish About the Actor?
Amazon assesses with moderate confidence that the operator has a Chinese nexus. It cites Simplified Chinese prompts, China Standard Time settings, and the selection of universities likely to be of strategic interest to China. That is Amazon’s attribution assessment, not a confirmed identity. The public evidence supports describing the operation without naming a specific state unit.
Amazon separately assesses with high confidence that the MONAI Label flaw was a zero-day at discovery. It found no corresponding fix, advisory, issue, or earlier database reference at the time. The maintainer advisory appeared later, on November 18. “Zero-day” describes the discovery and exploitation window, not the advisory’s current status.
Amazon reports that the operator later provided a tunnel into a second university network whose server ran the same vulnerable software. Amazon infers the operator had used the flaw to establish that tunnel. Its wording is an inference about the tunnel’s origin, while the first server’s root shell is a directly observed result. The article does not establish that every targeted university was compromised.
Brandefense Blackhole counted 76 dark-chat posts matching the broad term “Cairn” from August 23 through November 21, 2026. The measured unit is matching posts, and the term can refer to unrelated topics. This count is context for possible discussion only; it does not verify this campaign, identify an actor, or add affected institutions.
Our earlier Cairn framework intrusion report covers separate findings attributed to Unit 42 and GTIG. This investigation adds a specific MONAI Label product weakness and Amazon’s directly observed university campaign. Readers should not merge the timelines or treat separate research reports as one confirmed operation.
How Should a University Test Its Exposure?
Start with an inventory of research servers and medical imaging services, including systems run by individual labs. Compare hostnames, network routes, and installed MONAI Label package versions against the advisory. An external attack surface inventory can identify reachable services, while the local inventory confirms software and access settings.
For each MONAI Label instance, establish whether untrusted networks can reach it. Then verify whether authentication is enabled and which accounts hold the ANNOTATOR role. Check the process user and its filesystem write scope. The highest-priority cases combine a reachable service, an affected version, and broad process privileges.
Use the maintainer’s interim measures while waiting for a verified release. Restrict network access, enable authentication, remove the ANNOTATOR role from untrusted users, and run the server with limited filesystem privileges. These controls reduce possible access and impact, but they do not turn an affected package into a patched one.
- List every running MONAI Label instance and record its installed version.
- Confirm whether each service is reachable from untrusted networks.
- Check authentication settings and all ANNOTATOR role assignments.
- Review the service account and directories it can write.
- Preserve relevant application and host logs before changing a suspected system.
- Track the next maintainer release and verify the installed version after upgrading.
This checklist is our defensive workflow based on the maintainer advisory and Amazon’s account. It starts with exposure because a private lab service and a public endpoint have different immediate risk. It ends with version verification because the published advisory does not yet identify a patched package.
What Evidence Would Support an Incident Review?
Begin with the web, application, and host records for each exposed instance during the September campaign window and afterward. Look for unusual datastore upload activity followed by unexpected file changes or new application processes. Compare those events with authorized research jobs and maintenance. This is a hunting hypothesis derived from Amazon’s sequence, not a published detection signature.
On a suspected host, preserve logs and a copy of relevant filesystem metadata through the institution’s incident process. Review process ancestry, privilege level, and changes outside expected datastore directories. The incident response plan template offers a structure for ownership and evidence handling. Avoid declaring a compromise solely from an affected version or a reachability result.
Amazon says it worked with a government partner to notify affected organizations because it lacked a direct route to the non-AWS customers. A notified institution should reconcile the notice with its own host evidence. An institution without a notice still needs an exposure review if it runs the affected application. Absence of a notice is not an asset inventory.
The first decision is whether an instance is exposed and affected. The second is whether logs show the sequence Amazon described. The third is whether the eventual packaged fix has actually been installed and checked. Keeping those decisions distinct prevents unsuccessful agent attempts from being reported as successful breaches.
Frequently Asked Questions
Which MONAI Label versions are affected?
The November 18 maintainer advisory lists monailabel 0.8.5 and earlier. It says the fix is merged but does not identify a patched package release.
Did the agents compromise every targeted university?
Amazon observed one successful MONAI Label server compromise. It described other lengthy operations as unsuccessful and inferred, rather than directly proved, the origin of a later tunnel.
What should an exposed team do before a fixed release?
The maintainer advises restricting network access, enabling authentication, limiting ANNOTATOR access, and running the server with minimal filesystem privileges. Preserve evidence if compromise is suspected.
