New Today USA

AI Agents Found a Zero-Day Flaw and Took Over a University Server in 11 Minutes

For years, security experts have warned that AI would one day be able to hack on its own, finding weaknesses in software and exploiting them without a human telling it what to do. This week, Amazon published the first detailed account of that happening in the wild. Autonomous AI agents, left to their own devices, found a previously unknown flaw in a piece of software, used it to break into a university server, and took full control of the machine. The whole chain, from first look to root access, took about eleven minutes.

The disclosure came from Amazon Threat Intelligence, in a report published on 20 November on the AWS Security Blog by Amazon’s chief information security officer, CJ Moses. Its title said it plainly: a Chinese-nexus actor’s Cairn AI agents found a zero-day on their own. A zero-day is a flaw that nobody knew about before, which means there was no patch, no warning, and no defence waiting for it.

The story begins with a mistake by the attackers. On 10 September, Amazon found a server on the open internet running Cairn, an open-source framework for autonomous AI penetration testing, with no authentication on it at all. Anyone who found the server could see what the agents were doing, read their prompts, and watch their output. It was, in effect, a window into an active hacking campaign, left open by the people running it.

What Amazon saw through that window was a campaign that had begun two days earlier, on 8 September, and ran for about 48 hours. The operator had given the agents a list of target internet addresses, all belonging to universities in East Asia outside mainland China, and a single instruction in plain language: obtain shell access. The prompts were written in Simplified Chinese, and timestamps in the scripts the agents ran were set to China Standard Time. On that basis, and the choice of targets, Amazon assesses with moderate confidence that the actor is linked to China. It names no group.

The agents did what a human hacking team would do, only faster and without stopping. Against one university’s database server, they spent roughly sixteen hours and took 148 sequential actions without getting in. Against a university graphics computing cluster, they collected about eighty real usernames from faculty directories, lab pages, web archives, and academic papers, then tried more than thirty thousand credential combinations against its login systems, its virtual private network, its storage device, its admin dashboard, and its campus portal. They failed there, too. Those failures are as revealing as the success that followed, because they show the agents planning, trying, failing, and trying something else, exactly as a human team would, but without anyone choosing the next step.

The run that succeeded was against a server running MONAI Label, an open-source tool that research hospitals and universities use to label medical images and train AI models on them. The agents mapped the application’s programming interface, noticed that an upload endpoint did not properly confine the file name it was given, and tested it for a path traversal weakness. About eleven minutes after they started examining the application, they could write files anywhere on the server. In a later session, they used the same flaw to plant a small piece of code in a file that the Python programming language loads every time it starts, then called a training endpoint that started a new Python process. That process loaded their code, and gave them a shell running as root, the highest level of access on the machine. Amazon inferred that the operator then used the same flaw to compromise a server at a second university, tunnelling in through a proxy and directing the agents at other machines on that network.

The flaw is now tracked as CVE-2026-133968, with a severity score of 9.8 out of 10, close to the maximum. It affects all versions of MONAI Label up to and including 0.8.5, and it is especially dangerous because the software ships with authentication switched off by default. The maintainers merged a fix on 19 October, and the first patched release, version 0.8.6, was published on 22 November. Anyone running the software is urged to update now. Amazon says it is highly confident the flaw was a zero-day, because there was no fix, no advisory, and no prior record of it anywhere when the agents found it.

There is a detail in Amazon’s report that deserves attention, because it turns the usual story of a cyberattack upside down. Normally, investigators struggle to reconstruct what an attacker did, because the logs are incomplete or have been deleted. Here, the opposite was true. The attacker had left the AI framework’s own history open to anyone, so Amazon could describe the operation step by step. The most complete record of the attack belonged to the attacker, and it was the attacker’s own carelessness that made it public.

The universities involved have not been named, and the report makes no finding about whether patient or research data was taken. But the implications go well beyond two universities. For defenders, the lesson is that the time between a flaw being discovered and being exploited has collapsed. A human researcher might take days to turn a suspicion into a working exploit. These agents did it in minutes, and they did it without being told what to look for. For software makers, the lesson is blunter still. Shipping a medical imaging tool with authentication off by default, and with an upload feature that lets a caller choose where files go, is the kind of mistake that used to be survivable. In a world where autonomous agents are scanning the internet around the clock, it is not.

The report also raises a question that has no easy answer. Cairn is an open-source tool, built for legitimate security testing, and available to anyone. The same framework that a company’s own security team uses to test its defences can be pointed at someone else’s servers by an attacker on the other side of the world. That has always been true of hacking tools, but AI agents change the scale. One operator can now run dozens of agents at once, each working through the night, each learning from its failures, and each capable of finding something nobody knew was there. Eleven minutes, from first look to root access, is the new measure of how fast that can happen.

Exit mobile version