Amazon says Cairn AI agents found and exploited a MONAI Label zero-day
Amazon Threat Intelligence says a suspected China-linked actor pointed AI agents on the open-source Cairn platform at East Asian universities in September, and the agents found and exploited an unknown flaw in MONAI Label, a medical imaging tool, to get a root shell.
Warden, covering security, outages and the rules for tech
, 3 min read
11 minfrom mapping the API to arbitrary file writes
What to know
- Amazon says the agents went from mapping MONAI Label's API to writing arbitrary files on the server in about 11 minutes, then used that to get a root reverse shell.
- The flaw is a path traversal in MONAI Label 0.8.5 and earlier, tracked as GitHub advisory GHSA-qc22-rmgc-vg94 and rated critical; no CVE was listed.
- Amazon found the campaign because the attackers left their Cairn instance on the open internet without a login, exposing their prompts and the agents' output.
The operator's instruction was three words long: obtain shell access. Amazon says the AI agents did the rest, including finding a vulnerability nobody had reported yet.
Amazon Threat Intelligence published its account of the campaign this week, and BleepingComputer and others reported its findings on Friday. A suspected China-linked actor ran Cairn, an open-source autonomous penetration-testing framework, against universities in East Asia outside mainland China from September 8 to 10. One target ran MONAI Label, an open-source tool for labeling medical images. The agents mapped its API and had arbitrary file writes on the server about 11 minutes later. In a later session they used the flaw to drop a Python reverse shell into sitecustomize.py, then called a training endpoint that started a process running as root.
Kivorent News could not open Amazon's blog post or the GitHub advisory itself. The details here come from four reports of Amazon's findings that agree on the main points.
The flaw and the fix
The bug is a path traversal in MONAI Label's datastore upload endpoint, affecting version 0.8.5 and earlier. It is tracked as GitHub advisory GHSA-qc22-rmgc-vg94, rated critical, with no CVE listed. MONAI Label ships with authentication off by default, so an exposed server needs no login.
Patch status is the one point where the reports split. BleepingComputer, Mallory and ShiftDelete say the maintainers shipped a security update on November 18. WindowsForum, describing the advisory, says it lists no patched version and that the fix is a merged pull request waiting for the next release. If you run MONAI Label, check the advisory directly. Until a fixed release is installed, the steps named in that report are to turn authentication on, keep the server off untrusted networks and run it as an unprivileged user.
What the agents could not do
The failures are worth as much as the success. According to Mallory and ShiftDelete, the agents spent about 16 hours and 148 actions on a Redis host without getting in. They spent more than six hours trying over 30,000 credential pairs on a GPU server, and that failed too. The pivot into a second university network came through a proxy the human operator set up, and Mallory says a later step needed the operator's direct help.
So the system was fast where there was a fresh bug in an unauthenticated web app, and slow and noisy where there wasn't. That matches the FBI's warning last week about ARTEX, a similar framework, whose fastest intrusions reached privileged access in under three hours (our report). In both cases the edge is speed against apps already exposed to the internet.
Amazon rates the China link at moderate confidence, based on Simplified Chinese prompts, China Standard Time timestamps and the choice of targets. It rates the zero-day finding at high confidence, because no advisory, fix or public mention of the bug existed before. None of the reports names the language models behind the agents. Amazon told the victims, which were not AWS customers, through a government partner.
How did Amazon see all this? The attackers left their Cairn instance on a public server with no login, exposing their configuration, prompts and the agents' running output. The tools are new. The mistake that gave the operation away is old.
For background, see what a zero-day is and how to read a security advisory. We covered an earlier supply chain attack on AI developer tooling in the Microsoft agent-framework PyPI compromise.
Sources
- Hackers use autonomous AI tool Cairn to find zero-day, take over server, BleepingComputer, November 20, 2026
- AI Agents Found and Exploited MONAI Label Zero-Day in University Attacks, Amazon Says, Mallory, November 20, 2026
- Hackers Used Cairn AI Agents to Find a Zero-Day and Take Over a Server, ShiftDelete.Net, November 20, 2026
- Amazon: Cairn AI Agents Exploited MONAI Label Zero-Day to Gain Root Access, WindowsForum, November 20, 2026
AI agentsAmazonInternet, outages and security
Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.