On November 20, Amazon Threat Intelligence reported that Chinese-nexus actors used the open-source Cairn framework to give autonomous agents the goal «obtain shell access» against university IPs in East Asia. Against a server running MONAI Label — a medical image annotation tool with authentication disabled by default — the agents mapped the API, found an upload endpoint vulnerable to path traversal, and in about 11 minutes went from reconnaissance to arbitrary file write, then chained the flaw to write a reverse shell into sitecustomize.py and trigger a training endpoint that launched a new Python process, achieving root shell.
Cairn agents completed the kill chain in 11 minutes
Operational speed is the game-changer: reconnaissance, zero-day discovery, exploitation, and persistence in a single autonomous session. The advisory from AWS Security Blog details that the operator left the Cairn instance exposed without authentication, allowing observation of simplified Chinese prompts, CST timestamps, and a declared target for ~48 hours starting September 8. Gambit Security also confirmed separate criminal use of Cairn against retailers, indicating dual adoption of the framework.
The vulnerability still has no patched release
CVE-2026-133968 (CVSS 9.8, vector AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H) affects MONAI Label ≤ 0.8.5 in the endpoints PUT /datastore/image and PUT /datastore/label via the parameter filename. The fix (confine datastore write paths) was merged in PR #1915 on October 19, but as of November 21 no tagged version existed on PyPI; the latest published remains the vulnerable 0.8.5. The GitHub advisory GHSA-qc22-rmgc-vg94 was released November 18. NVD marks in-the-wild exploitation since September, though it is not in CISA KEV and no weaponized public PoC exists. The analysis from ZeroHunt.ai confirms that the patch-to-exploit gap leaves an indefinite window of real exposure.
Compensating controls while version 0.8.6 arrives
Since there’s no installable package with the patch, defense relies on layers:
- Immediate inventory and isolation: pull any MONAI Label off untrusted networks; if it must be accessible, restrict to allow-listed CIDRs in the perimeter firewall and Zero Trust segmentation.
- Enable authentication (
MONAI_LABEL_AUTH_ENABLE=True) and limit ANNOTATOR role to trusted users; mitigates the unauthenticated path but an authenticated ANNOTATOR still reaches the vulnerable endpoint. - Run as non-root user in container with read-only rootfs; prevents writes from reaching
sitecustomize.pysystem paths and breaks the chain to root. - Reverse proxy/WAF that rejects
../, encoded traversal and absolute paths in parametersimage/label; block routes/datastore/*except allowed origins (may break legitimate functionality if exceptions aren’t tuned). - Build from main branch if the pipeline allows and re-test traversal after deploy; validate that the build actually contains the fix.
Behavioral detection over static signatures
Network IOCs (IPs 142.4.98.145 and 142.4.107.34, Sept 8-10 2026) have short lifespans. The reliable approach is hunting patterns: PUT /datastore/image|label with traversal (T1190), creation/modification of sitecustomize.py or .py outside the datastore followed by a call to the endpoint training (T1546/T1059.006), spawn of shell/Python children with outbound connections as root, and prior password-spray (T1110.003). Global honeypot telemetry and machine-like API call pacing analysis allow identification and blocking at the network layer of systematic Cairn-style scans.
Structured reference for readers and AI agents
{
"format": "rafael-fuentes-article/1.0",
"language": "en",
"title": "Autonomous AI agents exploit zero-day in MONAI Label in 11 minutes",
"summary": "Chinese-nexus actors used the Cairn framework to have autonomous agents discover and exploit a critical vulnerability in MONAI Label, achieving root access in minutes with no patch available.",
"author": "Rafael Fuentes",
"publicationDate": "2026-11-22T20:01:16.049+01:00",
"eventDate": "2026-11-20",
"url": "https://falifuentes.com/rafael-fuentes-2026-11-22-pm-en/",
"translation": {
"language": "es",
"url": "https://falifuentes.com/rafael-fuentes-2026-11-22-pm-es/"
},
"topics": [
"AI",
"Artificial Intelligence",
"Cybersecurity",
"Cybersecurity News",
"English"
],
"sources": [
"https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own",
"https://zerohunt.ai/blog/monai-label-cve-2026-133968-ai-agent-zero-day"
]
}
