AI agents working for a suspected China-linked attacker discovered a previously unknown vulnerability on their own and used it to take full control of a university server, Amazon Threat Intelligence reported on Friday, November 20. Why it matters: earlier AI hacking campaigns mostly automated known techniques — this one involved the agents finding a brand-new flaw, which is a step up in what an unsupervised attack can do.
November 20, 2026
What happened
According to Amazon, the operator gave the agents a list of target IP addresses and one instruction, in variations of the phrase “obtain shell access.” From there, the agents planned, executed and adjusted their approach without the operator choosing each next step. The campaign ran for roughly 48 hours beginning September 8, 2026, against universities in East Asia outside mainland China.
Amazon found the operation on September 10, after the attackers left their instance of Cairn — an open-source penetration-testing framework that coordinates multiple AI agents — publicly accessible without authentication, exposing the operator’s prompts and the agents’ output. Amazon CISO CJ Moses said the agents “kept adapting for nearly a full day without human direction.”
The zero-day, found in about 11 minutes
The agents were not successful everywhere. An attempt on a university Redis server ran 148 sequential actions over about 16 hours and failed, and a separate effort against a university GPU server tried more than 30,000 credential combinations over six hours without getting in.
Then the agents turned to MONAI Label, an open-source medical-imaging AI application. They mapped its API, found an upload endpoint with a filename parameter, and identified a path-traversal vulnerability nobody had reported before. In about 11 minutes they went from mapping the API to writing files anywhere on the server. In a later session they used the flaw to plant a Python reverse shell in sitecustomize.py — a file Python loads on startup — and gained root, the server’s highest level of control. The operator then handed the agents a proxy into a second university network running the same vulnerable software.
Amazon assessed with high confidence that the flaw was a genuine zero-day, because there was no fix, issue, advisory or prior mention of it at the time. Amazon notified the MONAI Label maintainers, who released a security update and advisory on November 18; the advisory lists versions through 0.8.5 as affected. Amazon attributed the campaign to a China-linked actor with moderate confidence, citing Simplified Chinese prompts, script timestamps set to China Standard Time, and the choice of targets.
Why it matters
Cairn abuse is not new — Gambit Security documented it in a large-scale card-theft operation in September, and Google recently reported it in an attack on a U.S. aerospace organization. What is new here is the discovery step: the agents did not just run a known exploit faster, they found an unknown one. Moses’s warning to defenders follows from that speed — an agent that finds and exploits a flaw in minutes finishes before many security teams can manually triage the first alert, which is why Amazon is urging organizations to automate their response rather than approve each step by hand. For anyone running open-source AI tooling on a public server, the practical takeaway is unglamorous: patch MONAI Label, and check what your own agent frameworks are exposing to the internet — this campaign was only visible because the attackers left theirs open.
Frequently asked questions
What is Cairn?
Cairn is a legitimate open-source penetration-testing framework that coordinates several AI agents through multi-stage intrusion tasks. Like many security tools, it can be used by defenders for authorized testing or abused by attackers.
What software had the zero-day?
MONAI Label, an open-source application that uses AI to help researchers label medical images. The flaw was a path-traversal bug in an upload endpoint that let the agents write files outside the intended directory. A fix and advisory were released November 18.
Did the AI act completely on its own?
Not entirely. A human operator supplied the targets and the goal, and later provided a proxy into a second network. Between those inputs, Amazon says the agents chose and adapted their own steps, in one case persisting for about 16 hours without direction.
Sources: Amazon Threat Intelligence (via CJ Moses, AWS Security Blog), BleepingComputer, The Register

