MONAI Label: Cairn AI agents exploited path-traversal zero-day to get root on a university server (GHSA-qc22-rmgc-vg94)
Severity: critical · Patch status: partial · Since 2026-11-18 · CVE-2026-133968
Amazon Threat Intelligence (primary post dated 2026-11-20) reports that a Chinese-nexus actor, assessed with moderate confidence, pointed autonomous Cairn AI agents at IP addresses tied to universities in East Asia outside mainland China over roughly 48 hours beginning 2026-09-08. Redis, GPU-server and credential-spray attempts all failed. The agents then found that an upload filename parameter in an open-source medical imaging application allowed path traversal; in about 11 minutes they could write arbitrary files, planted a Python reverse shell in sitecustomize.py, and reached a root shell through a training endpoint. The actor then supplied a proxy into a second university network running the same software. Amazon does not name the product, but the GitHub advisory it matches (GHSA-qc22-rmgc-vg94, critical, CWE-22, no CVE) is for MONAI Label and credits Amazon's vulnerability-disclosure account as the original reporter. The fix was merged to the main branch on 2026-10-19, but no release containing it has been published (PyPI's latest monailabel is 0.8.5, from 2024), so today's remediation is a main-branch build with the fix or the advisory's workarounds: authentication on, the ANNOTATOR role withheld from untrusted users, the server kept off untrusted networks, and an unprivileged service account.
What to do
- No patched MONAI Label release exists yet: the GHSA lists no patched version and PyPI's latest is 0.8.5. Run a build of the MONAILabel main branch that includes commit 113527bf6 (PR #1915), or take the server offline until a release ships.
- Until then, set MONAI_LABEL_AUTH_ENABLE=True, withhold the ANNOTATOR role from untrusted users, keep the server off untrusted networks, and run it as an unprivileged user whose write access is limited to the datastore directory.
- Review MONAI Label and web server logs for PUT /datastore/image and PUT /datastore/label requests containing path-traversal sequences, and for unexpected file writes, changes to sitecustomize.py, or new Python startup files.
- Block and investigate the reported campaign IPs 142.4.98.145 and 142.4.107.34.
- Restrict outbound network access from application servers and segment research/GPU networks to limit reverse-shell and proxy pivoting.
- Require authentication and network controls on AI/agent tooling such as Cairn; do not expose management interfaces publicly.
- Monitor for anomalous API enumeration and upload requests against medical imaging endpoints.
Sources
All live incidents on Threat Watch