Cairn agents MONAI Label zero-day
A Chinese-nexus actor's Cairn AI agents found a MONAI Label zero-day and reached root on East Asian university servers
Posted · Spectrum Delta Research Team
Summary
A Chinese-nexus actor assessed by Amazon Threat Intelligence (B/1) set autonomous Cairn AI agents on East Asian universities with the single goal of obtaining shell access, and the agents found a path-traversal zero-day in the MONAI Label medical-imaging application, wrote a Python reverse shell to sitecustomize.py and reached a root shell in about 11 minutes. The agents ran roughly 48 hours of operations from September 8 to 10, 2026, including a 16-hour attempt against a hardened Redis service and more than 30,000 SSH credential combinations, before succeeding against MONAI Label; Amazon discloses the campaign in a November 20 report after finding the actor's Cairn server exposed without authentication. The maintainers published critical advisory GHSA-qc22-rmgc-vg94 on November 18 and NVD assigned CVE-2026-133968, but no fixed release has shipped and the fix exists only on the main branch. The actor later supplied a proxy into a second university network running the same software, which Amazon infers was compromised with the same zero-day. Attribution rests on Simplified Chinese prompts and China Standard Time timestamps and is assessed only with moderate confidence.
Breakdown
Countries
- East Asia
Targets
- Industry: Universities
Technical exposure
- MONAI Label through 0.8.5 with unvalidated datastore upload filename parameter
- MONAI Label deployments with authentication disabled, the default configuration
- Internet-facing Python services whose process user can write sitecustomize.py
- Internet-reachable SSH and unauthenticated Redis or Qdrant services
Adversary
Chinese-nexus threat actor · Amazon Threat Intelligence assesses with moderate confidence (B/1)
Procedures (5)
Credential spray against SSH and campus services with harvested usernames
Logs
Linux sshd auth log · Failed password and invalid-user events: one source testing more than 30,000 credential combinations against about 80 real usernames over six or more hours; OR auth failures on admin dashboards, NAS devices, VPN servers and campus portals
Detectors
- Elastic Security (prebuilt) · Potential Password Spraying Attack via SSH: bursts of failed SSH logins from a single source against many accounts in a short window
Sustained automated abuse of exposed Redis and Qdrant services
Logs
Redis server log · 148 sequential actions over 16 hours against an unauthenticated Redis instance: replication abuse to load code, scripting sandbox escape attempts, configuration file writes and memory corruption
- Qdrant and proxy logs · SSRF requests through a co-located Qdrant instance mapping the internal Docker network, followed by credential sprays against PostgreSQL, SSH and other internal services
Path traversal in MONAI Label upload writes files outside the datastore
- T1190Exploit Public-Facing Application
Logs
- Web server access log · PUT requests to /datastore/image or /datastore/label with an image or label parameter containing ../ traversal sequences; the agents went from API mapping to arbitrary file writes in about 11 minutes
Reverse shell payload written to sitecustomize.py
- T1546.018Event Triggered Execution: Python Startup Hooks
Logs
- EDR file events · Creation or modification of sitecustomize.py or usercustomize.py in site-packages, dist-packages or default sys.path directories by a web application process, here carrying a Python reverse-shell payload
Detectors
- Elastic Security (prebuilt) · Python Site or User Customize File Creation: file creation or modification of sitecustomize.py or usercustomize.py on Linux
- Falco · Modify Python Startup Hook File: open_write of sitecustomize or usercustomize modules in site-packages, dist-packages or default sys.path directoriesNot built-in
Python startup hook executes reverse shell as root after API trigger
- T1059.006Command and Scripting Interpreter: Python
Logs
- EDR process events · Web application spawning a new Python process that opens an outbound network connection, running as root after a training API call; process tree shows the interpreter importing the startup hook
- Web server access log · Training API endpoint call immediately preceding the Python process spawn, corroborating the trigger
Indicators of compromise
IPs (2)
- 142.4.98.145Cairn campaign infrastructure, records dated September 8-10 2026; hunt-only, perishable actor infrastructure assessed 2026-11-21
- 142.4.107.34Cairn campaign infrastructure, records dated September 8-10 2026; hunt-only, perishable actor infrastructure assessed 2026-11-21
CVEs (1)
- CVE-2026-133968MONAI Label datastore path traversal (GHSA-qc22-rmgc-vg94, critical), exploited as a zero-day by Cairn agents; NVD assigned November 20 2026; no fixed release as of 2026-11-21, fix on main branch only
Sources
- B/1AWS Security Blog: Chinese-nexus actor's Cairn AI agents found a zero-day on their own
- B/1Project MONAI security advisory: path traversal in datastore upload endpoints (GHSA-qc22-rmgc-vg94)
- B/1Elastic prebuilt detection rule: Python Site or User Customize File Creation
- B/1Falco incubating rules: Modify Python Startup Hook File
- B/1Elastic prebuilt detection rule: Potential Password Spraying Attack via SSH
Grades are source reliability (A–F) and information credibility (1–6).
Questions, detections or sightings of Cairn agents MONAI Label zero-day? Bring them to the community.