AI agents found and exploited a zero-day on their own — and took over a university server
Amazon says a suspected China-linked operator gave AI agents a single instruction — ‘obtain shell access’ — and the agents found a previously unknown flaw in a medical-imaging app, exploited it in about 11 minutes, and took root control of a university server.
- The campaign ran September 8–10, 2026, against universities in East Asia outside mainland China; Amazon disclosed it on November 20 in a post by its chief information security officer, CJ Moses.
- The agents used Cairn, an open-source AI penetration-testing framework, given only target IP addresses and the goal “obain shell access.”
- The zero-day: CVE-2026-133968, a path traversal flaw in MONAI Label (versions through 0.8.5), rated 9.8 out of 10.
- From mapping the app’s API to writing files anywhere on the server took about 11 minutes; a later session delivered a root shell.
- Amazon found the campaign on September 10 because the operator’s Cairn server was left publicly accessible with no login.
- A fix was released on November 18; Amazon attributes the campaign to a China-linked actor with moderate confidence.

A suspected China-linked operator handed a team of AI agents a single instruction — obtain shell access — and walked away. Over the next 48 hours the agents planned their own attacks, adapted when they failed, and eventually found a vulnerability nobody knew existed, exploited it, and took full control of a university server.
Amazon, which uncovered the campaign, disclosed it on November 20 in a post by its chief information security officer, CJ Moses. The company says the incident is the clearest evidence yet that autonomous offensive operations by AI agents are already happening — and that defenders need to start responding at machine speed.
One instruction, 48 hours, no human steering#
The campaign ran from September 8 to 10, 2026, and targeted universities in East Asia outside mainland China. The operator used Cairn, an open-source penetration-testing framework that coordinates several AI agents through multi-stage intrusions. The human supplied target IP addresses and variations on one goal — "obtain shell access" — and from there, Moses wrote, the agents "planned, executed, and adjusted their approach as they learned, without the operator choosing each next step."
Amazon's Threat Intelligence team found the campaign on September 10, and only because of an operator error: the Cairn instance running the attacks had been left publicly accessible with no authentication, exposing its prompts, settings, and logs to anyone who found it.
Eleven minutes from API map to root shell#
The agents were persistent, but not immediately successful. Over about 16 hours they ran 148 sequential actions against a university Redis server — abusing its replication feature, escaping its scripting sandbox, writing to configuration files — and failed. They tried a Log4Shell injection, a server-side request forgery attack against a Qdrant instance, and credential sprays against PostgreSQL and SSH. Against a university GPU server they tried more than 30,000 credential combinations over six hours. That failed too.

Then they probed MONAI Label, an open-source medical-imaging app researchers use to label scans. The agents identified the application's framework, mapped its API endpoints, and found an upload endpoint with a filename parameter. Testing it for path traversal — a flaw that lets an attacker write files outside the folder an app intends — worked. "In about 11 minutes, the agents went from mapping the application's API to arbitrarily writing files on the server," Amazon said.
In a later session, the agents used the same flaw to write a Python reverse shell into sitecustomize.py, a file Python loads automatically at startup, then triggered an API endpoint to run it — giving them a shell running as root, the machine's highest privilege. After that, the human operator stepped back in, supplying a proxy into a second university network running the same vulnerable software and directing the agents at more machines. Amazon inferred, rather than directly established, that the same flaw was used to compromise the machine providing that tunnel.
The flaw: a path traversal in MONAI Label#
The vulnerability is now tracked as CVE-2026-133968, with a severity score of 9.8 out of 10, and affects MONAI Label versions through 0.8.5. Amazon reported it to the project's maintainers, who merged a fix and published a security advisory on November 18.
Amazon assessed with high confidence that the flaw was a genuine zero-day — at the time the agents found it, there was no fix, no issue, no advisory, and no prior mention in any vulnerability database. Attribution is less certain: Amazon puts the China link at moderate confidence, based on prompts written in Simplified Chinese, script timestamps configured to China Standard Time, and targets it considers likely to be of strategic interest to China.

Cairn has shown up in attacks before: Gambit Security's threat-intelligence team documented a Chinese operator using the framework in at least 105 attacks. What makes this one different is the zero-day: the agents did not just reuse a known exploit — they discovered the vulnerability themselves.
What Amazon is telling defenders#
Moses's message to defenders is blunt: an agent that finds and exploits a zero-day in minutes "completes its activity faster than many security operations centers can manually triage an alert." His prescription: automated defenses — high-confidence detections that trigger automatic actions, with humans monitoring outcomes rather than approving each step.
Days earlier, Google's threat-intelligence group reported agents turning security patches into working exploits with no human in the loop. The Amazon case goes a step further: here the agents did the discovery too.
What to watch#
Three things matter next. First, whether other exposed Cairn instances — or copycat campaigns — surface now that the technique is public. Second, how quickly MONAI Label's users patch: the flaw affects every version through 0.8.5, and the software runs on GPU-equipped servers in hospitals, labs, and universities. Third, whether this incident sharpens the debate over agentic penetration-testing tools, built for authorized defense work but, as this campaign shows, needing only a target list and a one-line goal to run an attack.
Sources: The Register, 'China attackers deployed AI agents that autonomously found and exploited a 0-day to take over a university server,' November 20, 2026; BleepingComputer, 'Hackers use autonomous AI tool Cairn to find zero-day, take over server,' November 20, 2026; SecurityOnline.info, 'MONAI Label Zero-Day CVE-2026-133968 Exploited in the Wild,' November 21, 2026 (all reporting on Amazon Threat Intelligence's findings, disclosed by Amazon CISO CJ Moses).