MONAI Label zero-day CVE-2026-133968 now patched in 0.8.6
Part of the CVE-2026-133968 story · since 21 Nov
- Sat 21 Nov, 02:30 UTC. SecurityOnline, first report. Cairn AI agents exploit MONAI Label zero day for root access.
- Sun 22 Nov, 12:55 UTC. cybersixt, its first report on the story. MONAI Label zero-day CVE-2026-133968 now patched in 0.8.6. This article, 2nd of 2.
The developers of MONAI Label, an open-source server for AI-assisted medical image annotation, have released version 0.8.6, the first release that fixes CVE-2026-133968. That is the zero-day that Amazon says attackers' AI agents discovered on their own and used to get root on a university server. The release notes, published early on Sunday, call it a security release and say all MONAI Label server deployments should upgrade.
When Amazon Threat Intelligence disclosed the attack, the fix had been merged into the source code but there was no release to install.
What 0.8.6 fixes
The release fixes two path traversal flaws, both present in every version up to and including 0.8.5:
- CVE-2026-133968 (GHSA-qc22-rmgc-vg94, critical). The
PUT /datastore/imageandPUT /datastore/labelendpoints built the destination path from the image ID and label tag without confining it to the datastore, so a caller could write files anywhere the server process can write. - A second traversal in session uploads, with no CVE listed.
PUT /session/?uncompress=trueextracted uploaded tar archives without checking member names, and session IDs were used as paths without validation.
Authentication is off by default in MONAI Label, the release notes point out, so both flaws were reachable by any client that could connect to the server. The new version also rejects session archives that contain links or device files.
How the flaw was exploited
Amazon says a suspected Chinese-nexus operator, identified with moderate confidence, pointed AI agents built with the open-source Cairn framework at university targets and told them to obtain shell access. The agents mapped MONAI Label's API, found the upload path traversal and, within about 11 minutes, were writing files on the server. In a later session they wrote a Python reverse shell to sitecustomize.py, which Python imports at startup, then triggered a training endpoint to start a new Python process, which gave them a shell running as root. Amazon reconstructed the campaign, which targeted universities in East Asia outside mainland China over about 48 hours from September 8, from a Cairn instance the operator had left publicly accessible without authentication.
What to do
Upgrade with pip install -U monailabel==0.8.6 or docker pull projectmonai/monailabel:0.8.6. Version 0.8.6 requires Python 3.11 or newer. Whatever the version, the maintainers advise:
- do not expose the MONAI Label server to untrusted networks
- enable authentication and grant roles only to trusted users
- run the server as an unprivileged user whose write access is limited to the datastore
If a server running 0.8.5 or older was reachable from an untrusted network, the maintainers say it should be reviewed for unexpected file changes before it goes back into service.
What is not known
Neither Amazon nor the MONAI Label maintainers have said how many servers were exposed or compromised, and CISA has not added the flaw to its list of known exploited vulnerabilities.
How the story developed
2 articles from 2 outlets since 21 Nov. Still developing; the latest report was 4 hours ago. This was the 2nd report, 1 day after the first.
- Sat 21 Nov, 02:30 UTC. SecurityOnline, first report. Cairn AI agents exploit MONAI Label zero day for root access.
- Sun 22 Nov, 12:55 UTC. cybersixt, its first report on the story. MONAI Label zero-day CVE-2026-133968 now patched in 0.8.6. This article, 2nd of 2.
- CVE not in KEV
- No CVE
- Larger: an outlet’s first report
- Coverage over time
Show all 2 articles as a listHide the list
- Sat 21 NovDay 01
- Cairn AI agents exploit MONAI Label zero day for root access02:30 · SecurityOnline · First report
- Sun 22 NovDay 11
- MONAI Label zero-day CVE-2026-133968 now patched in 0.8.612:55 · cybersixt · This article, 2nd of 2
Sources
More reporting
- Cleverbridge data breach: stored payment card details copied22 Nov
- Nissan hack shuts dealer diagnostic system used for key fobs22 Nov
- OpenRouter cuts 3 AI providers over leaked customer prompts22 Nov
- Fluidkey hack: $1.6M stolen via session tokens in SigNoz logs22 Nov
- GoFundMe data breach: nonprofit donor data accessed via API21 Nov
Written by CyberSIXT from the sources above.How we reportCorrectionsRSS