Cairn AI Agents Found a MONAI Label Zero-Day in 11 Minutes, Amazon Says
Amazon Threat Intelligence says a suspected China-linked actor used Cairn, an AI agent platform built for penetration testing, to find an unknown path traversal flaw in the MONAI Label medical imaging server and exploit it within 11 minutes. The agents later gained root on a university server. The fix is merged, but no patched release had shipped by 21 November.
On this page
What Amazon says the attackers did
Amazon Threat Intelligence has described a campaign in which attackers handed most of the hacking work to AI agents, according to BleepingComputer's report on Amazon's findings. Amazon attributes the activity to a Chinese-nexus threat actor with moderate confidence, based on the language of the prompts it recovered, the timing of the activity, and the targets chosen. Those targets were mainly educational institutions in East Asia outside China, and the activity ran from 8 to 10 September 2026.
Amazon found the operation because the attackers left their own agent console open on the internet without authentication, which exposed their prompts and the agents' logs.
Cairn and the 11-minute zero-day
The tool at the centre of the case is Cairn, which BleepingComputer describes as a legitimate penetration testing platform that coordinates several AI agents and was built for research and defensive testing. The same tool has reportedly been misused in other campaigns, including one reported by Google against a US aerospace organisation.
Here the agents were pointed at MONAI Label, an open-source server used to label medical images for training AI models. According to the report, the agents went from mapping the application's API to exploiting a previously unknown flaw within 11 minutes. In a later session they wrote a Python reverse shell into a file that loads whenever Python starts, triggered it through an API call, and gained root access to the server. The operator then gave the agents a route into a second university network running the same software.
The agents did not succeed everywhere. A 16-hour attempt against a Redis server failed, and so did six hours of trying more than 30,000 password combinations on a university GPU server. None of the sources says whether any patient or research data was taken.
The MONAI Label flaw and its fix
The project's security advisory, published on 18 November, rates the bug critical. Two upload endpoints, for images and for labels, accepted a caller-supplied name and used it to build a file path without checking it, so a value such as a series of "../" steps could write files anywhere the server process had access. Versions up to and including 0.8.5 are affected, and no CVE had been assigned.
The default setup made this worse. Authentication is off by default in MONAI Label, so any client that can reach the server can exploit it; with authentication on, the ordinary annotator role is enough. The fix, which confines writes to the datastore folders and strips path parts from image names, was merged on 19 October. The advisory says it will ship in the next release, and the releases page still showed 0.8.5 as the latest version.
How to protect self-hosted AI tools
This case matters beyond hospitals and universities. Many people now run self-hosted AI tools such as labelling servers, model servers and notebook servers on local networks, and these are often left with default settings. The advisory's advice applies to all of them: turn authentication on, never expose the server to untrusted networks, and run it as an unprivileged user that can only write to its own data folder. Until a patched MONAI Label release appears, users can build from the main branch, which contains the fix.
Amazon's broader warning, as reported, is that AI agents can now find and use flaws faster than defenders can respond by hand, so organisations need more automated defences. Keeping internal tools off the open internet is the cheapest of those defences.