Amazon Says Suspected Chinese Attackers’ AI Agents Found a MONAI Label Zero-Day on Their Own
Amazon's threat team says AI agents run by a China-linked actor found and used an unpatched flaw in MONAI Label with no human choosing each step. What Amazon found, what it does not prove, and the checks that matter for the services the agents failed against.

Amazon says a China-linked attacker pointed a set of AI agents at university networks with a one-line goal, “obtain shell access”, and that the agents found a previously unknown flaw in an open-source medical imaging application and used it to take over a server. Amazon’s CISO, CJ Moses, published the account on Friday 20 November in a report on the AWS Security Blog. The software involved is MONAI Label, and Amazon says the maintainers have since released a fix and a security advisory.
The campaign itself dates from early September, and none of its successful steps need a human to have chosen them. That is the part worth reading closely, and it is also the part to treat with some care, because almost everything below comes from Amazon’s own telling of what its investigators found. We have not seen an independent analysis of the same server.
#What Amazon says it saw
On 10 September, Amazon’s threat intelligence team found a server running an instance of Cairn, an open-source framework for running AI penetration-testing agents, with no authentication in front of it. Because the operator left it open, Amazon could read the project settings, the operator’s prompts and the agents’ output. The prompts were in Simplified Chinese and the timestamps in the scripts were set to China Standard Time. On that basis, and because the targets were universities in East Asia outside mainland China, Amazon says it assesses “with moderate confidence” that the actor is Chinese-nexus. That is Amazon’s attribution, at a stated confidence level, and it is not an official statement from any government.
The operator supplied a list of IP addresses and a plain-language goal. According to Amazon, the agents then planned, tried things and adjusted over about 48 hours starting on 8 September, without the operator choosing each step. Cairn, Amazon notes, was built as a research tool for defenders.
Most of what the agents tried failed, and Amazon is explicit about that. Against a university Redis database that allowed access without a password, the agents spent roughly 16 hours and 148 actions on known techniques for taking over an open Redis server, then tried Log4Shell against another service on the same machine, then used a request-forgery weakness in a neighbouring Qdrant database to map the internal container network, then sprayed credentials at PostgreSQL, SSH and other services. None of it worked. A separate run of more than six hours against a university GPU server built a list of about 80 real usernames from faculty pages, lab sites, archived pages and papers, and tried more than 30,000 credential combinations. It never found a working login.
Then the agents met an application where the outcome was different. Amazon says they identified its framework, mapped its interface, found an upload feature with a filename parameter, and tested it for path traversal, the class of bug where a crafted filename lets a user write files outside the folder meant for uploads. It was vulnerable. By Amazon’s count, about 11 minutes passed between the agents mapping the application and being able to write arbitrary files to the server. In a later session they used the same flaw to get code running as the root user.
Amazon says the flaw matched no advisory, fix, issue or earlier mention in vulnerability databases for the project, and it assesses “with high confidence” that it was a zero-day. In a further session, after the agents had found it, the human operator supplied a network tunnel into a second university, which Amazon infers was reached by using the same flaw on another server running the same software.
#What this does and does not show
It shows that an operator can now hand a goal to agents and leave them running. The Register, reporting Amazon’s findings, quotes Moses saying the agents “kept adapting for nearly a full day without human direction”. The Register also notes an earlier report from Gambit Security that documented a Chinese operator using Cairn in at least 105 attacks and compromising, to varying degrees, at least 27 companies; Amazon links to the same research and says it shows the framework being adopted in both espionage and criminal operations.
It does not show that these agents are generally better than human attackers. By Amazon’s own account, they failed against two hardened-looking targets and succeeded against one application with a mistake in a single parameter. The number that matters is the 11 minutes, not the 48 hours. Nor does it tell us how many organisations were affected: Amazon says the victims were not its customers, so it worked with a government partner to notify them, and it names none. We would not read this as proof that defenders are suddenly outmatched. We read it as a clear sign that the time between a flaw existing and someone finding it is shrinking, and that automation that runs unattended is cheap.
#Who needs to check something
If you or a supplier run MONAI Label, which is used for annotating medical images and training models, treat this as an update to apply and a service to put behind authentication. Amazon says a fix and a GitHub Security Advisory exist; the advisory is GHSA-qc22-rmgc-vg94. We could not open that page from our systems, so we cannot give you the affected or fixed versions here, and we have not tried to guess them. Read the advisory, install the fixed release, and do not leave the service reachable from the internet. Amazon describes the agents reaching its web interface directly.
If you do not run it, the lessons are about the things that failed, and they are more useful than the zero-day. A Redis instance that accepted connections without a password was the most heavily attacked target, for about 16 hours; that it held out is no reason to leave it open. A database exposed to a container network, an SSH service answering credential guesses, and a campus portal with predictable usernames were all tried. Start with the unglamorous list: which of your database, admin and remote-access services answer from outside the building, which of them accept an unauthenticated connection, and which usernames can be guessed from your public website. Our SSH entry is a short refresher on what to lock down.
Amazon’s own advice to defenders is to put weight on high-confidence detections that trigger an automatic response, with people overseeing outcomes, because an agent that finds and uses a flaw in minutes will beat a person triaging each alert. We agree with the direction and would add a caveat of our own: automatic blocking of an address, an account or a service needs a quick, documented way back, or the first false positive will make the team switch it off.
#If you run a small business without an IT team
Unless you run research software, this story is unlikely to be about a system you own. The part that applies is old advice that has become more urgent: anything on the internet that asks for a password should have a long, unique one and a second check, and anything that does not need to be on the internet should not be. Ask whoever looks after your IT or hosting which services are open to the world. Our small business section has plain-English guides on account security.
#What we expect next
We expect more reports of agent-driven intrusions that describe the same shape: long, patient failure, then one quick success against something unpatched or unauthenticated. We also expect the arguments over tools like Cairn to continue, since the same framework that defenders use for testing is, by Amazon’s account, being run by attackers. What would prove us wrong is evidence that these runs needed more human steering than Amazon describes, for instance from an independent analysis of the exposed server. Amazon has not released that material, only the findings and two indicator addresses covering 8 to 10 September.
#What we do not know
We do not know which universities were affected, how many servers ran the vulnerable software, or whether the flaw was used elsewhere. We do not know the affected and fixed versions of MONAI Label, for the reason given above. And we have only Amazon’s word for how much of the campaign ran without human direction.
Evidence trail
Sources and verification
Primary documentation and external technical references used in this article.
About this page
Credited to
Dates
Published
Corrections
No corrections are listed for this page. Some review sweeps corrected pages without naming each one; those are in the corrections log.
More news
All news

Artificial IntelligenceMicrosoft Will Switch Claude On in Copilot for Tenants That Have Not Chosen, From 9 December

Was this useful?
Reports go to the site owner and are used to correct pages. Corrections are listed at /corrections.
Get new fixes by email.
New runbooks, configuration traps and fixes from KBY Technologies, about twice a week.
Comments
Add a thoughtful note on Amazon Says Suspected Chinese Attackers’ AI Agents Found a MONAI Label Zero-Day on Their Own. Comments are checked for spam and held for moderation before appearing.