⌁AI·CYBER·BRIEF▌
AI Threats7 min read

Amazon: China-linked AI agents found and exploited a zero-day in MONAI Label medical-imaging AI

Cairn AI agents went from mapping a MONAI Label server to writing files on it in about 11 minutes, Amazon says. No fixed release exists yet: lock the server down now.

Amazon Threat Intelligence says AI agents run by a suspected China-nexus operator found a previously unknown flaw in an open-source medical-imaging AI server, MONAI Label, and used it to gain root access, going from mapping the application to writing files on it in about 11 minutes. The flaw affects MONAI Label 0.8.5 and earlier, and no fixed release has shipped yet. If you run MONAI Label, take it off untrusted networks, switch on authentication and run it as an unprivileged user today.

What happened, in plain English

MONAI Label is a free tool that hospitals, universities and research labs use to label medical images (for example, marking a tumour on a scan) and to train AI models on them. It runs as a small web server that people and other programs talk to through an API (application programming interface, the set of web addresses a program exposes for others to call).

On November 20, 2026, Amazon's chief information security officer CJ Moses published an analysis of a campaign Amazon found in September. An operator gave Cairn, an open-source "autonomous penetration testing" tool, a list of target IP addresses and a goal: get shell access, meaning the ability to run commands on the machine. Cairn's AI agents then chose and adjusted their own methods.

On one target they found a zero-day (a flaw nobody, including the software's maintainers, knew about yet). MONAI Label's upload feature trusted a file name supplied by the caller. By crafting that name, the agents could save files outside the folder where images belong, anywhere on the disk the server could write to. They used that to plant code that the server later ran with full administrator ("root") rights.

An analogy: a mailroom that files every parcel wherever the label says, including "the CEO's desk drawer", without checking the address is inside the mailroom.

Amazon learned all of this because the operator's own Cairn control panel was on the internet without a password, exposing prompts and agent output.

Are you affected? What to do now

Most readers have nothing to do. This matters if your organisation, or a research group or supplier you work with, runs MONAI Label (the Python package monailabel). Hospitals, universities and AI research teams are the likeliest users.

According to the project's GitHub security advisory GHSA-qc22-rmgc-vg94, published November 18, 2026 and rated Critical:

  • Affected: monailabel 0.8.5 and earlier.
  • Patched release: none yet. The fix (pull request #1915) is merged in the source code and "will be included in the next MONAI Label release".
  • Default risk: authentication is off by default. With it on, an account with the ANNOTATOR role is enough to exploit the flaw.

Checklist

  1. Find it. Search your asset inventory, Python environments and container images for monailabel. Ask research departments and imaging-AI suppliers directly; these servers are often set up outside central IT.
  2. Take it off untrusted networks. The advisory's own advice is: "Do not expose the MONAI Label server to untrusted networks." Put it behind a VPN or an identity-aware proxy, or firewall it to known clients.
  3. Turn on authentication. Set MONAI_LABEL_AUTH_ENABLE=True and do not give the ANNOTATOR role to untrusted users (advisory wording).
  4. Run with least privilege. Run the server "as an unprivileged user with write access limited to the datastore directory" (advisory wording). Never as root.
  5. Look for compromise if it was exposed. Check for files written outside the datastore folder, especially Python files that load automatically when Python starts, unexpected outbound connections from the server, and new processes started by the training feature. If in doubt, rebuild the host and rotate any credentials stored on it.
  6. Check published indicators. Amazon lists two IP addresses seen in records dated September 8–10, 2026: 142[.]4.98.145 and 142[.]4.107.34. These are historical; treat a match as a reason to investigate, not proof.
  7. Watch for the release. Update as soon as a MONAI Label version above 0.8.5 ships with the fix.

The Hacker News reports that the flaw has been assigned CVE-2026-133968 with a 9.8 score. The GitHub advisory still showed no CVE when we read it, so treat that ID as reported, not yet confirmed by us.

The expert view

What the agents did, conceptually. Amazon describes a campaign of about 48 hours, starting September 8, 2026, against IP addresses "associated with universities in East Asia outside mainland China". The operator chose targets; the agents chose techniques. Much of what they tried failed: about 16 hours and 148 sequential actions against a Redis database server, and more than 30,000 password combinations against a GPU server using about 80 usernames built from faculty pages and papers. Neither worked.

The MONAI Label success followed a classic pattern. The agents identified the framework, mapped its API, found an upload parameter, tested it for path traversal (making a file path climb out of its intended folder) and confirmed they could write files anywhere. In a later session they turned that into code execution as root by placing code where the Python interpreter loads it automatically, then using an application feature that starts a new Python process. Amazon then saw the operator supply a proxy into a second university network running the same software and "infer[s] the actor had used the zero-day to compromise it". That second compromise is an inference, not a confirmed finding.

What is new. Path traversal is decades old, and the agents used no novel technique. What is new is documented, end-to-end autonomy: a machine found an unknown bug in niche software, confirmed it and weaponised it, with the human supplying only targets and a goal. Amazon assesses "with high confidence" that it was a zero-day, because no fix, issue, advisory or database entry existed at the time. The 11 minutes cover mapping the API through to arbitrary file write; the root shell came later.

Why it matters. Niche research software has long relied on obscurity: few attackers would spend days learning a medical-imaging labelling server. Agents remove that cost. Persistence is cheap too: 16 hours of failed attempts cost the operator almost nothing. Amazon's main advice is to "prioritize high-confidence detections that trigger automated response", with humans overseeing outcomes rather than approving every action.

Attribution. Amazon, not us, assesses "with moderate confidence" that the operator is a Chinese-nexus threat actor, based on Simplified Chinese prompts, China Standard Time timestamps in scripts and targeting of universities "likely of strategic interest to China". No group is named. Earlier, Unit 42 tied Cairn to a suspected China-nexus espionage campaign; Amazon's post does not link the two, and Cairn is open source, so shared tooling does not mean a shared operator. Gambit Security separately documented a financially motivated operator using Cairn against online retailers in September.

What is still unknown.

  • Which AI models or providers powered the agents. Amazon does not say.
  • Who the first victim was, what data the server held, and whether patient images were accessed.
  • How Amazon found the exposed Cairn panel.
  • Why no fixed release exists, when the source fix was merged on October 19 according to The Hacker News.

The broader pattern is consistent with Google's report that agents now turn patches into exploits and the ARTEX fork surge Cloudflare measured: the bottleneck in attacks is shifting from skill to the attacker's willingness to point agents at targets. For defenders, that makes unauthenticated, internet-facing AI and research servers the first thing to fix.

Official sources

Coverage:

Get the daily brief

AI + security signal by email: headlines, a two-line summary, a link. No noise, no spam.

How often