Amazon’s threat intelligence team did not set out to document an AI-driven intrusion. It found one because an operator left the control panel for an autonomous penetration-testing framework exposed to the open internet, with no password. The panel showed the prompts, the target list and the agents’ output.
What the team reconstructed was a roughly 48-hour campaign in September against universities in East Asia outside mainland China. The operator gave the agents a simple objective: obtain shell access. On one target, the agents found a previously unknown path-traversal flaw in MONAI Label, an open-source server for annotating medical images, and went from mapping its API to writing files on the host in about 11 minutes.
An accidental window into an agentic campaign
Cairn is an open-source framework that coordinates multiple AI agents to work toward a goal written in plain language. It was built as a research tool for defenders. The same design makes it useful to an attacker: the operator sets the target and the objective, and the agents choose the steps.
Amazon says it found the Cairn instance on September 10. The operator had written prompts in Simplified Chinese and configured script timestamps to China Standard Time. The visible records covered a campaign that began around September 8. Because the panel was public, Amazon could see the agents’ task settings, prompts and run history, a rare view of an intrusion that was still being planned and executed.
The failures are part of the story
Most of the campaign did not work. One operation against a Redis server at a university ran for about 16 hours and 148 actions. The agents tried replication abuse, scripting-sandbox escapes, configuration-file writes, memory corruption and Log4Shell injection. When those failed, they mapped a Docker network through an SSRF flaw in a co-located Qdrant service and sprayed credentials at PostgreSQL, SSH and other services. None of it produced a shell.
A second operation against a university GPU server lasted more than six hours. The agents built a list of about 80 real usernames from faculty pages, lab sites, archived pages and papers, then tested more than 30,000 credential combinations against SSH and several other services. They never found a working login.
That persistence is the point. The agents did not need a human to choose each next step, and they did not stop at the end of a workday. A defender who assumes an attacker will give up after a few failed attempts is assuming a human tempo.
The zero-day
The successful target ran MONAI Label, a server that research teams use to label scans and train imaging models. It is a separate project from the broader MONAI framework. Amazon says the agents identified the application, mapped its API endpoints and found an upload endpoint with a filename parameter. They tested the parameter for path traversal. It was vulnerable.
The flaw let an unauthenticated caller write files outside the intended storage folder. MONAI Label’s default configuration does not require a login, so any client that could reach the server could trigger the write. The server runs with the privileges of its service account, which limits the blast radius in theory but not in practice if the account can write to a code or startup location.
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 to spawn a new Python process. That process loaded the payload and connected back to the attacker as root.
Amazon rates the bug as a zero-day with high confidence because, at the time of discovery, there was no fix, issue, advisory or prior mention in vulnerability databases. The maintainers have since published a critical advisory covering version 0.8.5 and earlier. A fix was merged into the source code in October and is scheduled for the next release. As of this week, the advisory still listed no patched version, and the newest package release remained 0.8.5. The flaw is tracked as CVE-2026-133968 and rated 9.8 out of 10.
The pivot
After the agents had a foothold, the operator stepped in. Amazon says the operator gave the agents a proxy that tunneled into a second university network. The server behind the proxy was running the same vulnerable software, which suggests the actor had used the same flaw to compromise it and establish the tunnel. The agents were then told to target other machines inside that network.
Amazon attributes the campaign to an unidentified Chinese-nexus actor with moderate confidence. The evidence is the language of the prompts, the time zone in the scripts and the choice of targets. The company did not name a group and did not establish that a government was behind the operation.
What defenders should take from it
The first lesson is exposure. MONAI Label is a research tool, not a public web app. Servers that can be reached from the internet should be behind firewall rules, not just obscure URLs. Authentication should be on, and the ANNOTATOR role should be limited to people who need it.
The second is privilege. Run research services under a dedicated account with a narrow write scope. A path-traversal bug is dangerous because it turns a file write into a code-execution primitive. If the service cannot write to startup directories, the same bug is much less useful.
The third is detection. Amazon’s report describes a specific path: an unexpected Python startup file and an unusual training job. Teams running MONAI Label should check for both. They should also check whether a source build actually includes the October fix; a merged patch is not the same as a released package.
Amazon’s account is its own reconstruction. It has not been independently confirmed, and it does not show that patient scans or clinical systems were touched. The root shell was on one university server; another deployment may not give the same result. Those caveats matter. The broader finding does not depend on them: autonomous agents can now find and use a zero-day faster than many security teams can triage the first alert.
That is the uncomfortable part. The campaign’s failures show patience. Its success shows speed. Defenders will need both automation and human judgment to keep up.
Sources: AWS Security Blog; SecurityWeek; The Hacker News; BleepingComputer; VentureBeat.
Have a tip? Email [email protected] or use the contact form and choose Editorial. We do not promise anonymity.
Read next: ConnectWise ScreenConnect attacks hit nearly 50 organisations; SCKit malware: ScreenConnect attacks aim at MSP technicians; Optimizely’s signup flaw put malware on 178 companies’ sites. Browse the Security hub or all Security coverage.
