Cairn AI Agents Found MONAI Label Zero-Day in University Attacks
Amazon says a likely China-linked actor used the open-source Cairn AI pentesting framework to find and exploit a MONAI Label zero-day and gain root on university servers.

⚡ Why it matters
It is a documented case of autonomous AI agents finding and weaponising a new flaw in about 11 minutes, faster than many SOC teams can triage an alert.
Amazon Threat Intelligence has detailed a campaign in which a threat actor, assessed with moderate confidence to be China-nexus, used Cairn, an open-source framework that coordinates multiple AI agents for penetration testing, against educational institutions in East Asia between September 8 and 10, 2026.
How it was found
On September 10 researchers found the attacker's own Cairn instance exposed online without authentication, which revealed its operational history. Prompts were written in Simplified Chinese and timestamps were set to China Standard Time. The operator simply supplied target IP addresses and asked for shell access; the agents planned and adapted on their own.
The zero-day
The agents discovered a path traversal bug in MONAI Label, an open-source tool for AI-assisted labelling of medical images. It allowed arbitrary file writes, and the agents moved from mapping the API to exploitation in roughly 11 minutes. They later planted a Python reverse shell in sitecustomize.py and triggered an API endpoint to obtain root. The operator then pointed the tool at a second university running the same software.
Not every attempt worked: a 16-hour, 148-action effort against a Redis server failed, as did more than 30,000 credential attempts over six hours.
Amazon reported the flaw to the MONAI Label maintainers, who shipped a fix on November 18. No CVE identifier has been published in the reports so far.
Indicators
142.4.98.145 and 142.4.107.34 (observed September 8–10, 2026).
Recommendations
- Update MONAI Label to the November 18, 2026 release or later.
- Avoid exposing research AI applications to the internet without authentication.
- Hunt for connections from the listed IPs and for changes to sitecustomize.py.
- Invest in automated detection and response to match machine-speed attacks.
🔗 Sources
This article is original reporting by the Security World editorial team based on the sources listed above; direct quotes are kept short and marked with quotation marks. Corrections and feedback: info [at] securitywd.net.
Related stories
FBI Probes Apparent Breach of NCIC National Crime Information Database
The FBI is investigating an apparent hack of NCIC, the criminal justice database used by US law enforcement agencies. The scope of the incident is still unclear.
Rackspace Spot Breach: Leaked Credentials Exposed Kubernetes Secrets
Attackers used leaked credentials to access Rackspace Spot on Nov 17, creating pods and reading Kubernetes Secrets in some customer clusters. Rotate credentials now.
Microsoft Entra Connect Health API Flaw Exposed Admin Emails and Tenant Data
Microsoft says attackers abused an Entra Connect Health API bug to read admin emails and tenant metadata. A server-side fix is live; phishing risk remains.


