Skip to content
[ SIGNAL VERIFIED ] INDEPENDENT CYBERCRIME JOURNALISMCLASSIFIED // THE UNDERGROUND HAS A NEWSROOM / RSS
.COM
Latest News
ALL NEWS
Newsroom

Amazon Tracks Cairn AI Agents Exploiting a MONAI Label Zero-Day Against Universities

Amazon documented a suspected Chinese-nexus operator using Cairn autonomous agents to exploit an unpatched MONAI Label path-traversal bug. The maintainer has published a fix but no patched package release is listed yet.

Newsroom editorial artwork
PUBLISHED

IPCRIME | November 20, 2026. Amazon Threat Intelligence says a threat actor using the open-source Cairn multi-agent penetration-testing framework discovered and exploited a previously unknown file-writing weakness in MONAI Label, an AI-powered medical image annotation application used in research. The activity occurred in September but was publicly detailed by Amazon on November 20. In one observed session the agents moved from identifying a reachable application interface to writing files outside its intended storage area in about 11 minutes.

The report is significant because it describes a real campaign, not merely an AI demonstration in a laboratory. It is also important to keep the evidence precise: Amazon assessed a Chinese-nexus operator with moderate confidence, not a formally identified government unit or a proven state instruction. The target universities and exact countries have not been publicly named.

How investigators discovered the campaign

Amazon says that on September 10 its researchers discovered a publicly reachable, unauthenticated instance of Cairn. The operator had inadvertently exposed task history, agent outputs and prompts. Records placed the campaign approximately between September 8 and 10 and showed targeting of East Asian universities outside mainland China. The prompts used Simplified Chinese, with scripts using China Standard Time.

The operator set broad objectives rather than directing each individual action. Cairn coordinated tools and adjusted its approach as attempts succeeded or failed. Amazon documented a 16-hour sequence of 148 unsuccessful actions against a Redis environment and a separate attempt that tested over 30,000 password combinations without obtaining a valid login. Those failures matter: the research does not support calling all autonomous operations successful.

The MONAI Label weakness and the observed outcome

The successful branch involved a vulnerable MONAI Label server. The agents identified an upload interface where a caller-controlled file identifier could be used to construct a destination path without adequate boundary checks. The application could write outside its datastore directory. Amazon says a later session used the same flaw to obtain a privileged remote shell on one server. The actor subsequently provided a tunnel into another university environment using the same vulnerable application; Amazon infers prior compromise there, but the public record does not include the full logs or a count of data taken from either university.

MONAI Label’s security advisory GHSA-qc22-rmgc-vg94 describes the root cause as path traversal allowing arbitrary file writing. It lists monailabel 0.8.5 and earlier as affected and says a code fix is in project pull request #1915. As of the advisory reviewed today, the Patched versions field still says None: a merged or prepared code correction must not be confused with a published, upgradeable package release.

What server owners can verify now

  • Find actual MONAI Label deployments and check whether their web service can be reached from an untrusted network. An unused research workstation with no accessible service is not the same exposure as a publicly reachable server.
  • Follow the maintainers’ workaround: enable authentication, avoid granting the ANNOTATOR role to untrusted users, and restrict network access until a verified release containing the fix is available.
  • Run the service without administrative privileges and confine writable locations to the intended datastore. Amazon’s compromised server was reportedly running a process with excessive privileges.
  • Preserve relevant application and host logs. Investigate unexpected writes or process launches from the service rather than relying only on evidence of credential attacks.
  • Record the installed package build and check the project’s advisory for a concrete fixed version before describing the system as patched.

The broader question

Cairn is an authorized security research and assessment tool in legitimate contexts. A tool’s use does not identify its operator or establish that every user is a criminal. The observed risk is the speed with which an actor can coordinate analysis against an exposed weak service. That speed puts pressure on application developers to publish clear patch status and on administrators to maintain basic access boundaries.

Evidence and sources

Amazon Threat Intelligence — original November 20 investigation by CJ Moses; MONAI Label maintainers — GHSA-qc22-rmgc-vg94, affected versions and mitigations; MONAI Label pull request #1915.

IPCRIME analysis is based on those public records, not an independent intrusion reconstruction or hands-on exploit test. No university identity or verified exfiltration total has been established in the cited material.

Editorial note

IPCRIME distinguishes public records and documented findings from allegations, analysis and commentary. Report corrections through our contact page.

One response to “Amazon Tracks Cairn AI Agents Exploiting a MONAI Label Zero-Day Against Universities”