Skip to content
LIVE
The Executives BriefThe Executives Brief

Attackers' AI agents found a medical AI zero-day in 11 minutes. No patch yet

Amazon says the MONAI Label flaw let agents write files outside storage and then seize full server control, and the merged fix still has no release.

ByLama Al-RashidTechnology Correspondent, The Executives Brief
·4 min read
Attackers' AI agents found a medical AI zero-day in 11 minutes. No patch yet
Executive summary

Amazon said Friday that attackers' AI agents independently discovered and exploited a zero-day in MONAI Label, an open-source medical imaging AI application, writing files outside the intended storage directory in about 11 minutes and later gaining full control of a server. The fix has been merged into the source code but never released, leaving every deployment on version 0.8.5 or earlier exposed to a flaw that is exploitable by default and, in some cases, even with authentication enabled.

Amazon said Friday that attackers' AI agents independently discovered a zero-day in MONAI Label, an open-source medical imaging AI application, and used it to write files outside the application's intended storage directory in about 11 minutes. That is the headline number, and it is not the scariest part. The fix has been merged into the project's source code, but no patched release exists. MONAI Label's newest release on the Python Package Index is still 0.8.5, uploaded Nov. 25, 2024. The fix landed in the main branch on Oct. 19 in pull request #1915. So the vulnerability is public, the remedy is written, and every server running the current release is still exposed.

Amazon assesses with high confidence that the flaw was a zero-day because it found no prior fix, issue, advisory or mention in vulnerability databases at the time of discovery. MONAI Label's maintainers published a critical GitHub security advisory on Wednesday. It lists no patched version for 0.8.5 and earlier and says the fix "will be included in the next MONAI Label release." Translation for anyone running research infrastructure: there is no upgrade path today. There is only mitigation, and a clock.

How Amazon got the story is its own warning. The company found the attackers' Cairn server on Sept. 10 running without authentication and read the operator's prompts. Cairn is an open-source framework that coordinates AI agents for security testing. The campaign ran about 48 hours from Sept. 8 against universities in East Asia outside mainland China, and Amazon assesses a China nexus with moderate confidence. The operator gave the agents one goal: "obtain shell access." That is the whole brief. No step-by-step instructions, no list of specific exploits to try. Just an objective, handed to software that could search, attempt, and iterate on its own.

The agents did the rest. On MONAI Label, they discovered they could manipulate a file-upload feature to place files outside the application's intended storage area. In a later session, they exploited the same flaw to run malicious code and gain full control of the server. Two details make this worse than a typical bug. Authentication is disabled by default, the advisory says, meaning anyone who can reach the server can exploit the vulnerability. And even with authentication enabled, users assigned the ANNOTATOR role can still exploit it. The role you would hand to a graduate student doing labeling work is enough.

With no patched release available, administrators should first identify MONAI Label servers running version 0.8.5 or earlier. The advisory recommends three workarounds: enable authentication and restrict the ANNOTATOR role to trusted users; keep MONAI Label servers off untrusted networks; and run the software without administrator privileges while limiting its ability to write files outside its designated storage directory. Once a release containing the fix becomes available, administrators should upgrade. Because Amazon documented attackers exploiting the flaw, administrators should also check for signs their servers have already been compromised, including unexpected changes to application files. Enabling authentication now will not undo an earlier intrusion.

Step back and the structural problem is bigger than one project. Open-source medical imaging tools spread through exactly the channels that make them useful: research groups, university labs, hospital pilot programs, and graduate students who install a package and point it at a dataset. Those deployments rarely sit behind the same patch-management discipline as commercial clinical software. Historically, the gap between a fix being merged and a fix being released is measured in days. Here it has stretched from Oct. 19 to now, with 0.8.5 still the newest artifact on PyPI since Nov. 25, 2024. Every day that gap stays open, the advisory functions as a public map with no lock on the door.

The other shift is who did the finding. Amazon's account describes agents that were pointed at a target and, within roughly 11 minutes, produced a working path to write outside the sandbox, then escalated to full server control in a later session. For defenders, that compresses the window between a vulnerability existing and a vulnerability being weaponized. For attackers, it lowers the skill floor: the operator did not need to understand MONAI Label's upload handling, only to define success. The roughly 48-hour campaign against universities in East Asia is a reminder that the first targets of automated discovery are often institutions with valuable data and thin security teams, not large corporate security operations centers.

For executives and boards, the questions are concrete. Do you know which research-grade AI tools are running on your networks, and who owns patching them? When a maintainer says a fix is merged but unreleased, does anyone track that as an open exposure rather than a closed ticket? And does your threat model still assume adversaries need human expertise to find the next flaw? Amazon's report says otherwise. The fix for MONAI Label is written. It is just not shipped. Until it is, the workarounds and a compromise check are the only defenses available, and 11 minutes is the new baseline.

Executive ActionsLocked

This story's Key Insights and Take-aways are locked.

Create a free account to unlock Executive Actions for one credit.

Register to Unlock

Always free for Executives Club members. Join the Club

More in Technology