SIAgentNews.com
← All news

AI agents exploit medical imaging software to gain root access

Hackers used autonomous AI agents to find a zero-day flaw in MONAI Label, allowing them to compromise a university server.

· First reported by thehackernews.com

11 minutesTime taken for agents to reach the exploit point

The short version

  • Autonomous AI agents mapped the software interface and exploited a flaw in roughly 11 minutes.
  • The flaw allows attackers to write files outside the intended storage folder due to a lack of input validation.
  • The attack targeted MONAI Label 0.8.5 and earlier versions, which often run with no authentication enabled.
  • Investigators linked the campaign to a China-linked actor based on prompts in Simplified Chinese and local time settings.
  • The vulnerability is rated 9.8 out of 10 on the CVSS scale, indicating a critical security risk.

Why it matters

This event highlights a shift in cyber warfare where attackers can use autonomous agents to perform rapid reconnaissance and exploitation. It also underscores the risks of using open-source tools with default, insecure configurations in sensitive environments like healthcare.

The AI-driven intrusion

Suspected China-linked hackers recently utilized autonomous AI agents to identify and exploit a zero-day vulnerability in MONAI Label. This open-source server is commonly used by research teams to label medical scans and train imaging models. The attackers managed to gain root access to a university server by leveraging the software's inherent weaknesses.

Amazon Threat Intelligence discovered the breach after finding an exposed control panel for Cairn, an open-source framework for running AI agents. The panel revealed the operators' task settings, prompts, and run history for the agents. These agents were tasked with obtaining shell access, and they autonomously selected the methods to achieve that goal.

The agents mapped the application's interface and reached the point of writing files to the server in approximately 11 minutes. This rapid progression was achieved with very little step-by-step direction from the human operator. The human provided the high-level goal and targets, while the AI handled the technical execution of the penetration test.

How the flaw works

The vulnerability is a path-traversal flaw located in the parts of the server that handle uploaded images and labels. Specifically, the server takes a file identifier from a caller and appends it to the storage folder without checking the final result. This allows an attacker with no account to write files outside of the designated storage folder.

Because MONAI Label runs with the privileges of the account it is started under, an attacker cannot write to every part of the machine. However, if the server has a writable code or settings location, an attacker can use the flaw to place a file there. This creates a path to eventually running unauthorized code on the system.

The risk is heightened because authentication is turned off by default in the software. This means any client that can reach the server over a network can trigger the write operation. The flaw affects MONAI Label versions 0.8.5 and earlier, which are the versions currently listed in the project's release history.

Impact on research and security

The attack successfully resulted in a root shell on one university server. The agents achieved this by planting a file that Python runs automatically at startup and then triggering a training feature. This action started a new Python process that loaded the malicious file and connected back to the attacker as root.

While one server was confirmed to be compromised, the extent of the breach remains unclear. Amazon has not confirmed whether other MONAI Label systems were affected, though they believe a second university network might have been broken in the same way. No evidence currently suggests that patient scans, records, or clinical systems were touched during the incident.

The campaign also included several failed operations by the attackers. The success of the root access on one specific server should not be assumed as the outcome for every deployment. The severity of the risk depends heavily on how the specific server is run and what permissions its service account holds.

Patching and next steps

The maintainers of MONAI Label have merged a fix into the source code to address the path-traversal issue. This fix forces uploads to resolve within approved folders and rejects names that are empty or begin with a dot. However, as of the latest update, there is still no fixed version available for users to install.

The vulnerability has been assigned a critical score of 9.8 out of 10 in the CVE record. Because a merged source fix is not the same as a downloadable package, users cannot simply update to a new version. The project's security advisory currently lags in updating the CVE information, but the fix is intended for a future release.

Until a patched version is released, security experts recommend several defensive measures. Organizations should block untrusted network access to the servers, turn on the default authentication, and run the software under a limited account. It is also advised to check for unexpected Python startup files or unusual training jobs if a server was previously reachable.

What to watch

  • Whether a formal patched version of MONAI Label will be released soon.
  • How other organizations respond to the use of autonomous AI agents for penetration testing.
  • Whether more university systems using the software are identified as compromised.

Try it

If you use MONAI Label, check your server configuration to ensure authentication is enabled and that the service is not running with root privileges.

Sources and links

Read the original at thehackernews.com

First reported by thehackernews.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news