Amazon's threat team says an exposed server revealed how autonomous agents planned, failed, adapted and finally broke in.
· 4 min read
Amazon's threat intelligence team says a Chinese-nexus operator ran an open-source AI penetration-testing framework called Cairn against university servers in East Asia, and the AI agents found a previously unknown flaw in a medical imaging application and turned it into a root shell without a human choosing each step. The details, published by Amazon this week and summarised by Security Affairs and The Hacker News, are among the clearest accounts yet of AI agents discovering a real zero-day on their own.
An exposed server gave the game away
Amazon found the campaign by accident. On 10 September 2026, its team identified a publicly hosted Cairn instance that was running with no authentication, so anyone on the internet could read the project configurations, the operator's prompts and the AI's output. The prompts were written in Simplified Chinese, and the timestamps in the executed scripts were set to China Standard Time. The history showed a roughly 48-hour campaign, beginning on 8 September, against IP addresses linked to universities in East Asia outside mainland China.
The operator gave the agents a single instruction: obtain shell access. After that, according to Amazon's report as relayed by Security Affairs, no human chose the next step. The agents planned, executed and adjusted their own approach.
Where the agents failed
Most of the effort did not work. At one university the agents found a Redis server that required no password and spent about 16 hours on it, trying 148 different actions. Those included loading code through Redis replication, bypassing its protections, editing configuration files, corrupting memory and exploiting Log4Shell in a nearby service. When that failed, they abused a server-side request forgery flaw in a Qdrant server on the same network to probe internal systems, then tried different passwords against PostgreSQL, SSH and other services. None of it succeeded.
At another university the agents went after a GPU server for more than six hours. They gathered about 80 real usernames from university websites, faculty directories, research lab pages, archived pages and academic papers, then tried more than 30,000 username and password combinations against SSH, a VPN server and the university portal. That also failed. The point Amazon draws is persistence: the agents kept changing tactics without breaks or time limits.
Where they got in
The medical imaging target went differently. The Hacker News identified the software as MONAI Label, an open-source medical imaging server, with versions up to 0.8.5 affected in the default setup. The agents identified the application's framework, mapped its API endpoints and found an upload endpoint with a filename parameter. They tested the parameter for path traversal and found it vulnerable. Amazon says it took about 11 minutes to go from mapping the API to writing arbitrary files on the server.
A later session escalated that access. The agents wrote a Python reverse shell into a file called sitecustomize.py, which Python loads automatically on startup, then triggered a training endpoint that spawned a new process and imported the payload. The result was a root shell. Amazon assesses with high confidence that this was a genuine zero-day, because no fix, advisory or earlier mention existed for the affected project.
The human came back for the second stage
Once the bug had been proven, the human operator returned and gave the agents a proxy tunnel into a second university's network, where a server ran the same vulnerable software. Amazon infers that the actor had already used the same flaw to set up that tunnel, which would mean one AI-found bug unlocked a second network. Amazon notified the project's maintainers, who shipped a fix and a security advisory. The universities were not AWS customers, so Amazon routed its warning through a government partner. Separately, researchers at Gambit Security have documented a financially motivated actor using the same framework against online retailers, so the activity is not limited to espionage.
Part of a wider shift
The report lands after a run of warnings about autonomous attack tools. The FBI issued a flash alert on 13 November about ARTEX, another AI-driven penetration-testing framework that criminals have repurposed, and Google's Threat Intelligence Group said in a report dated 16 November that attackers had crossed a threshold from using chatbots for help to running agents that scan, exploit and adapt. Google said it had seen Cairn and PentAGI used in real attacks. Gadgets365 has covered related research this month, including OpenAI's claim that its pipeline built 948 working Linux root exploits and Google DeepMind's findings on legacy RSA.
What defenders can take from it
Speed is the central lesson. An agent that can go from API mapping to a file write in 11 minutes can outrun a team investigating a single alert, and these agents kept working for 16 hours straight on targets they never cracked. Amazon's advice, in its own security blog post, is for organisations to prioritise high-confidence detections that trigger an automated response, with humans overseeing outcomes rather than reviewing every individual action. Teams that run MONAI Label should check their version and configuration against the maintainers' advisory.
Published 22 November 2026. Gadgets365 will update this article as more information becomes available.
Editor
Pravin KankanePravin Kankane is the editor of Gadgets365, an independent technology publication based in Pune, India. He covers smartphones, AI, Apple, Android and consumer tech launches in India.