Last month I wrote about Cairn, the open-source pentest framework criminals were using to run AI hacking agents. Amazon Threat Intelligence just published the next chapter: a Cairn operator's agents found a previously unknown bug in a medical imaging AI app at a university and turned it into a root shell, with the hard part taking about 11 minutes.
It's a real milestone. It's also a very fixable one, so let's stay practical.
What Amazon Found
From CJ Moses's November 20 post on the AWS Security Blog:
- On September 10, Amazon found an unauthenticated, internet-exposed Cairn instance. Yes, again. The operators keep leaving the control panel open.
- The campaign ran about 48 hours from September 8, against IP ranges at East Asian universities.
- The target was a public MONAI Label server, an open-source tool for labeling medical images for AI training.
- The agents mapped the API, found an upload endpoint with a filename parameter, and found it was vulnerable to path traversal. Amazon: "In about 11 minutes, the agents went from mapping the application's API to arbitrarily writing files on the server."
- In a later session they wrote a Python reverse shell into
sitecustomize.pyand called a training endpoint to spawn a Python process, which got them a root shell. - Amazon assesses "with high confidence" it was a zero-day, and "with moderate confidence" the actor is China-nexus, based on Simplified Chinese prompts, China Standard Time timestamps and the targeting.
- The operator then gave the agents a proxy into a second university running the same software, and Amazon infers it was compromised the same way.
The Part That Didn't Work
This is the part I'd want every CISO to read. The same agents spent about 16 hours and 148 actions against a Redis-based database node, trying replication abuse, sandbox escapes, Log4Shell injections, an SSRF pivot and credential sprays. Amazon: "None of these attempts succeeded." Against a GPU server they tested more than 30,000 credential combinations over six hours. "They never found a working login."
So the agents are persistent, cheap and fast at finding the one broken thing. They are not magic. Patched services with decent credentials held up for a full day of automated attempts. The one that fell was an internet-exposed research app with authentication off by default.
The Fix
Amazon notified the MONAI Label maintainers, who published GitHub advisory GHSA-qc22-rmgc-vg94. Per the Cloud Security Alliance's write-up of the advisory, it's a critical path traversal in the datastore image and label upload endpoints, affecting monailabel 0.8.5 and earlier, with auth disabled by default. The code fix was merged in October, but as of November 21 CSA reported PyPI still listed 0.8.5 as the latest release, and one outlet reported a 0.8.6 release on November 22. Check the advisory for the fixed version rather than assuming pip install -U got it.
Disclosure: I work at Zscaler. The defense here isn't a product pitch, it's mostly hygiene.
What To Do
- Get research and lab apps off the open internet. MONAI Label, Jupyter, model-serving and annotation tools should sit behind authenticated access, not a public IP. Universities and hospitals especially.
- Turn on authentication (
MONAI_LABEL_AUTH_ENABLE) and limit who has annotator rights. - Run these apps as an unprivileged user, not root. Root is what turned a file write into a full takeover.
- Hunt for persistence: check
sitecustomize.pyand.pthfiles in Python environments, and look for../in upload requests to the datastore endpoints. - Restrict outbound traffic from research servers so a reverse shell has nowhere to call.
- Automate the response side. Amazon's own advice is to prioritize "high-confidence detections that trigger automated response, with humans overseeing outcomes." If the attack takes 11 minutes, a ticket queue isn't the defense.
Bottom Line
AI agents found a real zero-day and used it. They also failed for a full day against systems that were patched and properly authenticated. Exposure is still the whole game: if it isn't reachable, the agent can't map it.
Sources: AWS Security Blog: Chinese-nexus actor's Cairn AI agents found a zero-day on their own, Cloud Security Alliance research note, MONAI Label advisory GHSA-qc22-rmgc-vg94, Gridinsoft