What happened
A path-traversal flaw in MONAI Label upload endpoints joins a caller-controlled ID onto storage paths without checking the result stays inside the datastore folder. Amazon Threat Intelligence documented a China-nexus operation in which agents from the open-source Cairn autonomous-pentest framework (whose operator panel was left unauthenticated on the internet) mapped the API and achieved arbitrary file write in about eleven minutes, then chained the bug to a root shell in a later session. Published 2026-11-20, CVSS 9.8 CRITICAL, and the CVE record notes exploitation in the wild starting September 2026.
Why it matters
This is the first documented real-world incident where an autonomous AI-agent framework independently discovered a zero-day, weaponized it, and escalated to root with little step-by-step human direction — machine-speed compromise of ML research infrastructure, and a template for agent-driven intrusion against exposed ML/imaging servers.
Attack vector
Unauthenticated network request to the PUT /datastore/image (and /datastore/label) upload endpoints supplies a caller-controlled identifier that is appended to the storage path and written to disk without confinement (directory traversal). In the observed campaign the path traversal was used to write a Python startup hook (sitecustomize.py) and then trigger a training endpoint that spawned a Python process which loaded the payload and connected back as root.
Affected systems
MONAI Label (monailabel) <= 0.8.5
Mitigation
No patched release exists yet (advisory lists 'none'; newest PyPI release is still 0.8.5). Mitigate: block untrusted network access to the server, enable authentication (MONAI_LABEL_AUTH_ENABLE), run under a least-privilege account and restrict writable folders, and only trust a source build that includes the Oct 19 PR #1915 fix. Advisory GHSA-qc22-rmgc-vg94.