AI agents exploited a zero-day vulnerability in MONAI Label
Amazon reports observing autonomous attacks carried out with Cairn, an AI pentesting framework, against universities in East Asia. The agents reportedly found a zero-day vulnerability in MONAI Label and gained root access.
Amazon states that operators linked to China used Cairn, an open-source AI-driven offensive auditing framework, to attack universities in East Asia and discover a zero-day vulnerability in MONAI Label, a medical image labeling tool. The critical point is not only the discovery of the vulnerability: the agents reportedly chained reconnaissance, file writing, and code execution without human supervision at any stage. The campaign, observed in early September 2026 according to Amazon Threat Intelligence, mainly demonstrates how quickly AI can turn a breach into an exploitable intrusion.
An attack that advanced on its own, or almost
The Cairn server spotted by Amazon was publicly accessible, without authentication. Logs exposed the prompts, configurations, and outputs of the agent, written in Simplified Chinese with timestamps set to China Standard Time. Based on this, Amazon says it reconstructed a campaign targeting university IPs in East Asia outside mainland China over about 48 hours, starting September 8, 2026.
The modus operandi differs from classic intrusions. Initially, the agent tested several exposed services, including Redis, PostgreSQL, SSH, and a Qdrant server, without lasting success. On another target, it collected nearly 80 real credentials from university sites, then attempted over 30,000 username and password combinations. Nothing poetic, but a pace many monitoring centers dislike very much.
MONAI Label, the vulnerability that shifts the balance of power
The success came from an open-source medical image annotation application, MONAI Label, used in research and imaging workflows. Amazon indicates that the agents mapped the API, identified a file upload parameter vulnerable to path traversal, then achieved arbitrary file write in about 11 minutes. The vulnerability, tracked as CVE-2026-133968 with a CVSS 3.1 score of 9.8, allowed escaping the expected directory and writing where the process had write permissions.
The mechanism is classic but formidable. In practice, such a flaw can be used to drop an SSH key, overwrite a configuration file, or place a Python file triggered automatically at startup. This is precisely what was done here: the agent wrote a sitecustomize.py, then triggered a training API to launch a new Python process that imported the payload and opened a reverse shell with root privileges. The magic, once again, was mainly in the automation.
Why this vulnerability was exploitable so quickly
MONAI Label has a dangerous blind spot: authentication is disabled by default, and even when enabled, an account with the ANNOTATOR role is sufficient to reach certain write points. The maintainers then fixed the issue by confining write paths and adding input validation to the HTTP interface. The fix is available in the main branch via a path confinement pull request but had not yet been released in a packaged version at the time of the sources.
Amazon says it notified the maintainers, who prepared the patch and security advisory. The company also indicated it notified the affected organizations via a government partner, without naming the universities or country. According to sources, the same access was then used to pivot to a second university network via a tunneled proxy. This kind of chain gives a clear idea of the problem: a single vulnerability, and the rest follows very quickly.
What defenders should watch for
Mitigation recommendations converge on three simple measures. First, enable authentication and do not assign the ANNOTATOR role to untrusted users. Next, avoid exposing MONAI Label to uncontrolled networks. Finally, run the service with a non-privileged account and limit its write permissions to the designated storage directory only.
On detection, the most useful signals remain concrete: PUT requests to /datastore/image or /datastore/label with directory traversal sequences, unexpected file writes like sitecustomize.py, then creation of a child Python process that opens an unusual outgoing connection. Amazon also mentions two IP addresses found in campaign logs, 142.4.98.145 and 142.4.107.34, to be used for retrospective filtering over the targeted periods. In short, the right response is not to wait for a perfect alert; it is to block quickly when an out-of-scope write attempt appears.
Key points
- Amazon places the campaign in early September 2026.
- Cairn operated without authentication, with visible prompts as evidence.
- MONAI Label served as an entry point to root access.
- The successful attack took about 11 minutes after API mapping.
- Two IPs are cited by Amazon: 142.4.98.145 and 142.4.107.34.
- The MONAI Label patch is in the main branch, not in a release.
By the numbers
- 48 hours — approximate duration of the campaign observed by Amazon, starting September 8, 2026.
- 11 minutes — time between API mapping and arbitrary file write.
- 148 actions — sequence of unsuccessful attempts against a first university server.
- 30,000+ combinations — credential attempts against SSH and other services.
- CVSS 9.8 — severity published for CVE-2026-133968.