Autonomous AI agents exploited a zero-day in medical imaging software
Amazon Threat Intelligence reports that suspected China-linked hackers used autonomous AI agents to discover and exploit a critical path-traversal flaw in MONAI Label, gaining root access to universit
Suspected hackers linked to China used autonomous artificial intelligence agents to discover a previously unknown vulnerability in MONAI Label, an open-source server for medical imaging. Amazon Threat Intelligence reported this week that the attackers leveraged these automated tools to gain root access to a university server by exploiting the flaw in September.
What happened
The target was MONAI Label, a server application used by research teams to label medical scans and train imaging models. It is distinct from the broader MONAI framework. The vulnerability, tracked as CVE-2026-133968, is a path-traversal flaw affecting version 0.8.5 and earlier when running in the default configuration. In this setup, authentication is disabled, allowing any network client to interact with the server without credentials.
Amazon reconstructed the intrusion after discovering the attacker’s own control panel, known as Cairn, exposed on the internet without password protection. Cairn is an open-source framework designed to run AI agents as automated penetration testers. The exposed panel provided investigators with the agents' task settings, prompts, and execution history, offering a rare look into how autonomous systems conduct cyber operations.
Over a period of approximately 48 hours in September, the AI agents mapped the application’s interface and identified the vulnerability. According to Amazon’s report, the agents reached the point of writing files to the server in roughly 11 minutes. The operators provided high-level goals, such as obtaining shell access, but the agents selected the specific methods and execution steps themselves with minimal direction.
How it works
The core weakness is a path-traversal issue in the file storage mechanism. When a user uploads an image or label, the server takes a file identifier provided by the caller and appends it to the designated storage folder path. The system then writes the file to disk without verifying that the resulting path remains within the intended directory. This allows an attacker to specify a path that escapes the storage folder, such as using directory traversal sequences like ../.
Because the server runs with the privileges of its host account, the attacker can write files to any location writable by that account. While this does not immediately grant full system control, it creates a foothold. If the attacker can write to a location where code or configuration files are stored, they can execute arbitrary commands. In this case, the attackers planted a file that Python executes automatically at startup. They then triggered a training feature, which started a new Python process. This process loaded the malicious file and established a reverse connection to the attacker, granting them a root shell.
Key details
- The vulnerability affects MONAI Label versions 0.8.5 and earlier in the default no-login configuration.
- CVE-2026-133968 has been assigned a critical severity score of 9.8 out of 10.
- Maintainers merged a source code fix on October 19, but no patched installable version has been released yet.
- The attack was conducted using Cairn, an open-source framework for autonomous AI penetration testing.
- Investigators attribute the campaign to a China-linked actor with moderate confidence based on language and time zone artifacts.
- The initial file write occurred in about 11 minutes, but privilege escalation to root required a separate, subsequent session.
Why it matters
This incident highlights a shift in how vulnerabilities are discovered and exploited. Traditionally, finding a zero-day flaw requires significant manual effort or sophisticated fuzzing infrastructure. Here, autonomous agents performed the reconnaissance and exploitation with little human intervention. For developers and security engineers, this means that exposed services face a higher volume of automated, intelligent probing. The speed at which the agents identified the flaw underscores the risk of leaving development or research tools exposed to the public internet, even if they seem obscure.
The situation is complicated by the lack of a ready-to-install patch. While the maintainers have fixed the code in their repository, users cannot simply update via a package manager. This gap between source code fixes and distributed releases leaves systems vulnerable for an indefinite period. Teams relying on open-source medical imaging tools must now actively monitor source repositories and build from source if they need immediate protection, rather than waiting for official release notes.
Furthermore, the attack vector demonstrates the danger of default configurations. MONAI Label ships with authentication disabled to facilitate ease of use in research environments. However, when deployed on networks with any external exposure, this default becomes a critical liability. The combination of a path-traversal bug and a permissive default setting allowed the attackers to move from remote file write to full root access, illustrating how configuration choices can amplify technical vulnerabilities.
What you can do
- Block untrusted network access to MONAI Label servers using firewall rules and routing policies, rather than relying on obscurity.
- Enable authentication on all MONAI Label instances, as it is turned off by default.
- Run the server under a limited, non-root user account and restrict the directories it can write to.
- Verify that any source builds you compile include the fix merged on October 19, as no patched binary release exists yet.
- Inspect servers for unexpected Python startup files and unusual training job logs if they were previously exposed to the internet.
- Monitor for updates from the MONAI project regarding the next official release that will include the security patch.



