Eleven Minutes From First Look to a Working Exploit: A Timeline of How Attackers’ AI Agents Found a Zero-Day in Medical-Imaging Software
Amazon says a suspected China-linked operator pointed an open-source AI agent framework called Cairn at university servers, told it to get shell access, and watched it discover an unknown flaw in the MONAI Label imaging server and eventually reach root. Here is how the campaign unfolded, and where the fix stands.
Security teams have been warned for months that attackers would start handing the hands-on work of hacking to AI agents. Amazon’s threat intelligence team has now published one of the most detailed accounts yet of that happening: a campaign in which the agents, given little more than a list of targets and a goal, found a brand-new vulnerability on their own and used it to take over a server. The chronology below is assembled from The Hacker News’ breakdown of the incident, with additional detail from The Record and BleepingComputer.
The setting
The target software was MONAI Label, an open-source server that research groups use to annotate medical scans and train imaging models. It is a separate program from the wider MONAI framework. The tool doing the attacking was Cairn, a legitimate open-source framework built to coordinate multiple AI agents as automated penetration testers. Before this case, Gambit Security had documented Cairn being abused in a card-theft operation against online retailers, and Google has since reported seeing it in intrusions at U.S. aerospace companies.
How it unfolded
September 8 to 10: the agents go to work
Over roughly 48 hours, an operator aimed Cairn at internet addresses belonging to universities in East Asia, outside mainland China, and gave the agents a plain-language objective: obtain shell access. According to Amazon, the operator picked the targets and the goal, and the agents chose their own methods, planning and adjusting as they went.
The misses
Much of the campaign failed. One effort ran to 148 actions across 16 hours without success, including an attempt to use the old Log4Shell bug; BleepingComputer reports that this effort was aimed at a Redis server. Another spent more than six hours trying over 30,000 username and password combinations, built around 80 real usernames, against a university GPU server and never got in.
The hit: about 11 minutes
The one success came when the agents identified MONAI Label on a target and began mapping its programming interface. Roughly 11 minutes later, they had found and used a path-traversal flaw in the upload functions that store images and labels. Because the server joins a caller-supplied file name to its storage folder without checking where the result ends up, the agents could write files outside that folder. Authentication is switched off by default, so no account was needed.
A later session: root
The root shell did not come in those first 11 minutes. In a separate session, the agents used the same flaw to plant a file that Python loads automatically every time it starts, then set off a model-training feature that launched a new Python process. That process picked up the planted file and connected back to the attacker with root privileges on that one university server.
A second university
The operator later gave the agents a tunnel into a second university network and told them to go after other machines there. The server at the other end of that tunnel was running the same vulnerable software, and Amazon infers it was broken into the same way, though it has not confirmed that compromise.
September 10: the attacker’s control panel is found
Amazon’s investigation began when it came across the operator’s own Cairn instance sitting on the internet with no password. That exposed the agents’ prompts, task settings and run history, which is how investigators were able to reconstruct the operation step by step. Prompts written in Simplified Chinese, China Standard Time settings and the choice of targets led Amazon to attribute the activity, with moderate confidence, to a China-nexus actor it has not tied to any named group. Amazon says it worked with a government partner to notify the affected organizations, which it has not named.
October 19: a fix in the source code
After Amazon reported the bug, the MONAI Label maintainers confirmed it and merged a correction that forces uploads to stay inside approved folders and rejects empty file names or names starting with a dot.
This week: advisory, CVE and a warning
The project published a security advisory rating the flaw critical and covering version 0.8.5 and earlier. It is now tracked as CVE-2026-133968 with a score of 9.8 out of 10. The advisory says the correction will ship in a future release, and The Hacker News notes that the newest installable release is still 0.8.5, from 2024. In its report, Amazon warns that agents working at this speed can finish their work before many security teams have even triaged the first alert.
What the report does not say
Amazon’s account is its own reconstruction and has not been independently verified. It does not show that patient scans or records were taken, that clinical systems were touched, or that the attacker got beyond the university targets described. And the root outcome on one server depends on how that server was set up; a MONAI Label installation running under a restricted account would not necessarily give up the same access.
What this means for your company
Few small or mid-size businesses run medical imaging research servers, but almost every one runs some specialist application that sits on the network with its default settings: a lab tool, a scanner portal, a line-of-business server a vendor installed years ago. The lesson of this campaign is that obscure software is no longer protected by being obscure. An agent can read an unfamiliar interface and probe it for mistakes far faster than a person, and it does not get tired. Make a list of every application that listens on your network, especially anything reachable from the internet. Turn on authentication wherever it is optional, run services under limited accounts rather than administrator or root, and put anything that does not need to be public behind a firewall or VPN. Then make sure someone, or some automated tool, is actually watching for unusual activity around the clock, because a human reviewing alerts the next morning may simply be too late.
If you are not sure which of your applications are exposed or running with more privilege than they need, we can help you map them and tighten them before someone else’s AI agents go looking. Reach out to Layer 7 Data Solutions to discuss how we’ve helped companies similar to yours.
