Actively exploited Security news
Cairn AI agents found a MONAI Label zero-day on their own and took over university servers
On 20 November 2026 Amazon Threat Intelligence published an analysis of attacks on East Asian universities in which AI agents running the Cairn framework found a previously unknown flaw in MONAI Label, an open-source medical imaging application, and used it to get root without step-by-step human direction. The fix is in the code but not yet in a release. If you run MONAI Label, put it behind authentication now.
Key points
- On 10 September 2026 Amazon found a publicly reachable, unauthenticated Cairn instance complete with its operation history. The operator gave the AI agents only IP addresses and the goal "obtain shell access".
- The agents spent 16 hours and 148 actions on Redis and more than 6 hours testing over 30,000 credential combinations, both without success. They succeeded against MONAI Label, going from API mapping to arbitrary file write in about 11 minutes.
- The bug is a path traversal in file uploads (advisory GHSA-qc22-rmgc-vg94, now CVE-2026-133968, CVSS 3.1 9.8). MONAI Label ships with authentication turned off.
- Amazon assesses with moderate confidence that the actor is China-nexus (Simplified Chinese prompts, China Standard Time timestamps, choice of targets) and published two IP addresses.
What happened
On 10 September 2026 Amazon Threat Intelligence came across a publicly reachable instance of Cairn, an open-source framework in which autonomous AI agents carry out penetration tests. It had no authentication, so Amazon could read the full history of its operations. Starting on 8 September, it spent roughly 48 hours attacking IP addresses belonging to East Asian universities outside mainland China. Amazon published its analysis on 20 November 2026.
The operator gave the agents nothing but addresses and the goal "obtain shell access". The agents planned their own steps:
- Unauthenticated Redis: about 16 hours and 148 actions (replication abuse, sandbox escape, config writes, memory corruption), plus a Log4Shell attempt and password sprays against PostgreSQL and SSH. All failed.
- GPU server: more than 6 hours, around 80 usernames gathered from public sources and over 30,000 credential combinations against SSH, a VPN, storage and a campus portal. All failed.
- MONAI Label: in this medical image annotation app the agents found a path traversal in the filename parameter of a file upload. Amazon says they went from mapping the API to writing arbitrary files in about 11 minutes. In a later session they wrote a reverse shell into
sitecustomize.pyand ran it as root by triggering a model training endpoint.
The operator then handed the agents a proxy into a second university's network, which Amazon infers was breached with the same bug. Amazon assesses with high confidence that the flaw was a zero-day: no advisory or prior mention existed.
Technical details
According to advisory GHSA-qc22-rmgc-vg94 (rated critical), the server does not validate the image ID passed to PUT /datastore/image, and the label-saving endpoint has the same flaw. A crafted ID writes the uploaded file anywhere the server process can write. Authentication is off by default, so anyone who can reach the server over the network can exploit it. Versions 0.8.5 and earlier are affected. The fix is merged as PR #1915 and, per the advisory, will ship in the next release. As of 21 Nov 2026, PyPI still offers only version 0.8.5, released in November 2024.
Since 20 Nov 2026 the flaw has been tracked as CVE-2026-133968 (assigned by MITRE). The record scores it CVSS 3.1 9.8 (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H), classifies it as CWE-24 and says it has been exploited since September 2026. NVD has not analysed it yet.
How to tell if you were hit
Amazon's indicators (records from 8-10 September 2026, defanged):
142[.]4[.]98[.]145
142[.]4[.]107[.]34
- Search firewall and web server logs for both addresses. You can paste log exports into our IOC extractor.
- On MONAI Label hosts, look for
sitecustomize.pyfiles in Python directories and unexpected files outside the datastore. - In access logs, look for
PUT /datastore/requests containing../.
What to do
- Do not expose MONAI Label to the internet. Turn on authentication or put it behind a VPN or an authenticating reverse proxy.
- Apply the fix from PR #1915, or update as soon as the new release is out.
- Plan for speed: Amazon warns that agents can exploit a bug within minutes and work without breaks, so high-confidence detections should trigger automated response.
For how a different group used the same framework against chipmakers, see Cairn: autonomous AI agents hit chipmakers and defense.
Timeline
Update
- 21 Nov 2026 - the flaw is now CVE-2026-133968 (CVSS 3.1 9.8, published on CVE.org on 20 Nov). No fixed release is on PyPI yet; only the source-code fix is available.
- 20 Nov 2026 - first version of this article.
FAQ
Is this the first time AI found and exploited an unknown bug by itself?
No. On 16 November Google described a similar case, covered in our article on AI attacks without a human in the loop. What makes Amazon's write-up unusual is the full operation history, failures included: the agents spend most of their time hitting walls and succeed only where there is a real weakness.
I use the MONAI library. Am I affected?
The advisory covers the MONAI Label server application (the monailabel package) up to version 0.8.5. The core MONAI training library is not listed in it.
Which organizations were hit?
Amazon names only universities in East Asia outside mainland China and has not identified them. Because the victims were not AWS customers, AWS worked with a government partner to notify them. Anyone running MONAI Label should check whether it is reachable without authentication.
Sources
- AWS Security Blog - Chinese-nexus actor's Cairn AI agents found a zero-day on their own (CJ Moses, 20 Nov 2026)
- GitHub - Project-MONAI/MONAILabel, GHSA-qc22-rmgc-vg94 (18 Nov 2026)
- CVE.org - CVE-2026-133968 record (published 20 Nov 2026)
- Unit 42 - analysis of the CL-STA-1238 Cairn campaign (12 Nov 2026)
Spotted a mistake? Let us know and we will correct it openly.