ResearchHot

An AI Agent Found a Zero-Day and Owned a University Server in 11 Minutes

An operator gave the agents one goal and no method. They found an undocumented flaw in medical imaging software and owned the server, after 148 failed actions somewhere else.

Emily CarterEmily Carter
Heat: 1,520
A stack of agent nodes linked by a path to a server rack, with a keyhole and a breach mark on the rack

Dek: An operator gave the agents one goal and no method. They found an undocumented flaw in medical imaging software and owned the server, after 148 failed actions somewhere else.

The 11 minutes from API map to shell on MONAI Label

The instruction was four words long: obtain shell access. Everything after that was left to the agents, according to Amazon Threat Intelligence's writeup published on Friday 20 November.

What followed reads like a penetration test report, because that's what the tool was built to produce. The agents identified the application, mapped its API endpoints, spotted an upload endpoint with a filename parameter, tested it for path traversal and confirmed the flaw. From API mapping to writing files on the server took about 11 minutes.

The software was MONAI Label, an open-source medical imaging app that uses AI to help researchers label scans. The bug was a zero-day, meaning nobody had documented it before the agents hit it. Path traversal is the trick of feeding a file path that escapes the folder it was meant to stay inside.

148 actions against Redis that went nowhere first

Persistence came before success, and the failures are the more useful half of the story.

Target

Effort

Result

Redis server

about 16 hours, 148 sequential actions

failed

University GPU server

more than 6 hours, 30,000-plus credential attempts

failed

Log4Shell, Qdrant SSRF, PostgreSQL and SSH sprays

separate attempts

failed

MONAI Label upload endpoint

about 11 minutes from API map to file write

worked

The Redis attempt included abusing replication to load unauthorised code, escaping a scripting sandbox, writing configuration files and corrupting memory. None of it landed. Sixteen hours, and nothing.

The GPU server got its own campaign. Over six hours, the agents compiled 80 real usernames from the school's public websites, papers and documents, then tried more than 30,000 credential combinations. That failed too.

How the Cairn agents turned a filename into root

The successful chain is short enough to describe in a handful of moves, which is the point.

An upload endpoint accepted a filename. The filename carried a path traversal. In a later session the agents wrote a Python reverse shell into sitecustomize.py, a file Python loads on startup, so the shell ran the next time anything started the interpreter. Then an API call turned that foothold into root on the server.

After the zero-day worked, the human operator handed the agents a proxy into a second university network running the same vulnerable software and pointed them at more machines. That detail matters for anyone tempted to read the story as software acting alone.

What Cairn is, and what it isn't

Cairn is an open-source penetration-testing framework that coordinates multiple AI agents. It was built for research and defensive security work, and Amazon's report treats it as a legitimate tool that got abused.

Not malware.

So where does the line between assisted intrusion and autonomous hacking actually sit? The report draws it at the goal. A human picked the target and the objective, and the agents picked every step in between, exploit included.

It also isn't new to this kind of use. Gambit Security documented a Chinese operator running Cairn for at least 105 attacks and compromising at least 27 companies in September, part of a card-theft campaign that leaned on two other open-source toolkits for running agents. Google has reported Cairn in an attack against a US aerospace organisation.

Amazon's China attribution sits at moderate confidence

The campaign ran for roughly 48 hours from 8 September and targeted universities in East Asia outside mainland China.

Amazon attributes it to an actor with links to China, at moderate confidence. The evidence it cites is circumstantial but specific: prompts written in Simplified Chinese, timestamps set to China Standard Time, and target selection that fits a regional pattern.

Two addresses were tied to the 8 to 10 September window, 142.4.98.145 and 142.4.107.34. Treat those as historical indicators from that week, not as a blocklist to paste into a firewall today.

The break in the case was the attackers' own Cairn instance

Amazon found the campaign in the least cinematic way possible. The attackers left their own Cairn instance hosted on the public internet without authentication, and it exposed their prompts and session logs.

That's how the defenders got the sequence above in such detail, including the exact wording of the operator's instruction.

The patch, and what defenders should take from Cairn

Amazon reported the flaw to the MONAI Label maintainers, who shipped a security update and an advisory on 18 November. If your lab or hospital runs MONAI Label, that's the update to install.

Universities sit at the awkward end of the risk map. Public staff directories, shared GPU clusters and research software installed by graduate students make for a soft target, and the agents read all of that the way a patient attacker would.

Amazon's own recommendation is to lean on automated detection and patching rather than manual investigation, because agents iterate faster than an incident response team can read tickets. Two inventories matter more than another awareness session: which services face the internet, and how fast a patch actually reaches production.

The same week, ITPro reported that AI had surfaced another decade-old flaw, this time in the nginx web server. One case proves little. A pattern of finding old bugs at machine speed is the thing to plan for, and the plan starts with the list of what you're exposing.

Share This Story

Sources

Related AI News

Concentric spiral curves crossed by a vertical line with bright points pinned along it
Research

An Anthropic Researcher Says Claude Proved the Riemann Hypothesis. Stanford Answered Within Hours

A researcher says he and Claude produced a proof of the hardest problem in mathematics, and the code compiles. Verification is a different job, and it has barely started.

Heat: 1,860
A funnel carrying small token blocks from one model box to another, with a shield plate blocking part of the flow
Research

OpenAI Cut Reasoning Theft From 54% to 0.2%. The Rest Is an Industry Problem.

OpenAI published the before-and-after numbers for its latest safeguards and they look like a fix. The same post explains why the fix has to be repeated every week, by everyone.

Heat: 1,420
A green chip grid with one highlighted cell and a jagged signal line beside the words Blackwell fingerprints
Research

Blackwell fingerprints in six Chinese models: what a checkpoint can prove

Dek: PyTorch fills billions of empty weights with random numbers, and the GPU doing the filling leaves a mark in the finished file. Someone checked six Chinese checkpoints and found newer silicon than the rules allow.

Heat: 1,250
Radio tower sending signal arcs over a grid of connected devices
Research

Nvidia's open source 5G and 6G toolkits passed a million downloads

Nvidia spent a year giving away the software researchers use to prototype wireless networks. The download counter passed a million, and the more useful signal is what a handful of labs built with it.

Heat: 1,275