Amazon Threat Intelligence described a case on November 20, 2026 in which AI agents found a zero-day and exploited it on their own. An attacker that Amazon links with moderate confidence to a Chinese-nexus actor used the open-source penetration testing framework Cairn against universities in East Asia. The agents were given only IP addresses and a goal: obtain shell access. Against a medical imaging AI application, they went from mapping its API to writing arbitrary files on the server in about eleven minutes. The affected software is MONAI Label. No fixed release has been published yet.
How Amazon found the Cairn instance
According to the report on the AWS Security Blog, Amazon came across a server on September 10, 2026 that hosted a publicly reachable Cairn instance without authentication. Because the operator had left it open, the platform’s entire operational history was visible: project configurations, prompts and the AI’s output. The prompts were written in Simplified Chinese, and timestamps in executed scripts were set to China Standard Time.
The campaign targeted IP addresses associated with universities in East Asia outside mainland China over roughly 48 hours starting September 8, 2026. Amazon says Cairn was built as a tool for defenders. You give it a starting point and a goal in plain language, and the framework coordinates several AI agents that look for a way to get there. That ease of use, Amazon notes, is exactly what makes it attractive to attackers.
16 hours of persistence without success
The report covers the failures as well as the hit. The most extensive attempt targeted a Redis database at one university that was reachable without authentication. Over about 16 hours and 148 sequential actions, the agents worked through a long list of known techniques: abusing replication to load unauthorized code, escaping the scripting sandbox, writing configuration files and corrupting memory. They also tried a Log4Shell injection against another service on the same system.
When that failed, the agents used a server-side request forgery flaw in a co-located Qdrant instance to map the internal Docker network and launched credential attacks against PostgreSQL, SSH and other services. None of it worked.
A second operation ran for more than six hours against a university GPU server. The agents built a list of about 80 real usernames from public sources such as faculty directories, lab websites, archived pages and academic papers. They then tested more than 30,000 username and password combinations against SSH and against other systems on the university network, including an admin dashboard, a network storage device, a VPN server and the campus portal. They never found a working login.
AI agents find a zero-day in eleven minutes
The attack on a target running an open-source medical imaging AI application went differently. The agents identified the framework, mapped the API endpoints and found an upload endpoint with a filename parameter. They tested it for path traversal, and the parameter was vulnerable. Amazon says about eleven minutes passed between mapping the API and writing arbitrary files on the server.
In a later session, the agents used the same flaw to write a Python reverse shell into sitecustomize.py, a file Python imports automatically when the interpreter starts. They then triggered a training endpoint in the application to spawn a new Python process. That process loaded the payload and opened a reverse shell running as root.
At the time of discovery, the flaw had no advisory, fix, issue or mention in vulnerability databases. Amazon therefore assesses with high confidence that it was a zero-day. In another session, the human operator stepped in and gave the agents a proxy into the network of a second university whose server ran the same software. Amazon infers that the attacker had already exploited the flaw there to set up the tunnel.
The flaw in MONAI Label
Amazon’s post does not name the application, but it points to the maintainers’ GitHub advisory. Advisory GHSA-qc22-rmgc-vg94, published November 18, 2026, covers MONAI Label, a tool from the MONAI project for annotating medical imaging data. It credits AWS’s security team as the original reporter. The key points:
- All versions up to and including 0.8.5 are affected, and the advisory lists no patched version yet.
- The
PUT /datastore/imageendpoint takes theimageparameter as the image ID without validation. A value such as../../../tmp/xwrites the uploaded file to any location the MONAI Label process can write to.PUT /datastore/labelis affected the same way. - Authentication is disabled by default (
MONAI_LABEL_AUTH_ENABLE=False). In the default configuration, any client that can reach the server can exploit the flaw. With authentication enabled, the ANNOTATOR role is enough. - The project rates the flaw critical (CWE-22). The advisory lists no CVE ID.
The fix is in pull request #1915 and is due to ship with the next MONAI Label release. On PyPI, version 0.8.5, published in November 2024, was still the latest release when this article was written on November 21. Amazon’s wording that the maintainers released a fix therefore refers to merged code, not an installable release.
What operators should do now
Until a new version is available, the advisory recommends these steps:
- Enable authentication (
MONAI_LABEL_AUTH_ENABLE=True) and do not grant the ANNOTATOR role to untrusted users. - Do not expose MONAI Label servers to untrusted networks.
- Run the service as an unprivileged user that can only write to the datastore directory.
Anyone running MONAI Label should also check for unexpected files such as a modified sitecustomize.py, new entries in .ssh/authorized_keys or new cron jobs. Amazon also published two IP addresses from September 8 to 10, 2026 as indicators: 142.4.98.145 and 142.4.107.34. The affected universities were not AWS customers. Amazon says it worked with a government partner to notify them.
The takeaway: persistence over brilliance
What stands out in this case is less a particularly clever technique than the combination of speed and persistence. A path traversal flaw in an upload parameter is a classic. What is new is that agents discovered it in a niche application without human guidance and immediately turned it into code execution as root. At the same time, the failed attempts show that well-secured systems can withstand 16 hours of automated attacks.
The case fits patterns we have covered recently, such as exploit_agent turning patches into exploits automatically and AI agents used to breach more than 100 companies. Amazon says Gambit Security has also documented Cairn use, in that case by a financially motivated group targeting online retailers.
Amazon’s advice to defenders is to rely on high-confidence detections that trigger automated responses, because agents work faster than many security teams can triage alerts by hand. In practice, that mainly means research and ML servers belong in the asset inventory and behind authentication just like VPN gateways and web servers. Specialized AI tools in particular often run on the network with default settings, and as MONAI Label shows, those are not always safe.




