
Amazon Threat Intelligence reports a China-linked operator used Cairn agents to find and exploit an unknown flaw in a medical imaging tool. The patch status is disputed across reports.
Amazon Threat Intelligence says AI agents run by a China-linked operator found and exploited a previously unknown flaw in a medical imaging tool in about 11 minutes, with no human writing the exploit. The report, published around November 20, 2026, describes a September 2026 campaign against East Asian university servers and is one of the clearest public cases of an AI agent framework doing zero-day discovery on its own. It also contains a patch-status conflict that defenders should resolve before relying on any one account.
What Amazon reports happened
Amazon found the activity after locating an exposed, unauthenticated instance of Cairn, an open-source AI penetration testing framework that coordinates several agents. The operator's prompts and agent output were visible on that instance, which is how the campaign was reconstructed. SecurityWeek, reporting on November 21, says the campaign lasted roughly 48 hours; BleepingComputer dates it September 8 to 10.
The target software was MONAI Label, an open-source tool for AI-assisted medical image annotation. The agents found a path traversal bug in a file upload function. Amazon's own wording, as quoted by SecurityWeek: "In about 11 minutes, the agents went from mapping the application's API to arbitrarily writing files on the server." They then planted a Python reverse shell in a file that runs at interpreter startup, triggered an API endpoint and obtained a root shell. Amazon says it assesses with high confidence that the flaw was a zero-day.
From there the operator used a proxy to reach a second university. Amazon infers the actor had compromised that server with the same zero-day to set up the tunnel, noting the server "was running the same vulnerable software."
The failures matter as much as the success
The same campaign shows where the agents hit limits. BleepingComputer reports a 16-hour, 148-action attempt to take control of a Redis server that failed, and more than 30,000 credential combinations tried over six hours against a university GPU server, also without success. One successful chain, then, sat alongside long unsuccessful grinds. That is consistent with what offensive teams already know: automation is cheap, so it can afford to fail often.
“In about 11 minutes, the agents went from mapping the application's API to arbitrarily writing files on the server”

Attribution and its confidence
Amazon attributes the activity to an unidentified Chinese-nexus actor with moderate confidence. BleepingComputer lists the basis as prompt language, activity timestamps and target choice; targets were mainly East Asian educational institutions outside mainland China. Moderate confidence is not a government attribution, and the operator is not named. Cairn is a legitimate security tool, so the story is about misuse of a dual-use framework, not about a purpose-built attack tool.
The patch status is unclear, so check the advisory
Sources disagree. BleepingComputer says maintainers released a patch on November 18, 2026. SecurityWeek quotes the advisory, GHSA-qc22-rmgc-vg94, as saying "A fix has been created and will be included in the next release," and lists 0.8.5 and earlier as affected. SecurityWeek also notes authentication is disabled by default in MONAI Label, so anyone who can reach the server over the network can exploit it. One security site names a CVE identifier and a score, but we found no second source for it and are not repeating it. We could not open the advisory page itself this run, so treat any claim about a shipped fix as unconfirmed until you check the maintainers' page.
The practical advice is plain: do not expose MONAI Label servers to untrusted networks, turn authentication on, and watch for the next release. The indicators of compromise in Amazon's report include two IP addresses that BleepingComputer prints in defanged form.
Where this fits in the larger trend
This is not the first Amazon report on agentic offense. Our tracker of the ARTEX hacking agent campaigns logs earlier cases of open-source agent frameworks being run by criminal and state-linked operators, including exposed consoles like the one that gave this campaign away. BleepingComputer adds that Gambit Security documented Cairn abuse in a credit card theft operation in September and that Google reported Cairn being used against a US aerospace organization.
What is new is the discovery step. Using an agent to scan for known flaws is old news. An agent that maps an unfamiliar API and finds a write primitive on its own, in minutes, compresses the time a defender has between a bug existing and a bug being used. It does not mean every system is now at risk from this technique; the flaw was a fairly classic path traversal in a niche tool with no authentication by default.
What we can and cannot say
We can say Amazon reports the 11 minutes, the zero-day assessment and the moderate-confidence China link. We cannot independently verify any of it: the Amazon blog page did not render in our fetch, so we relied on two outlets quoting it and on Amazon's own wording as they reproduced it. We have not seen the operator's prompts, and Amazon says the failed attempts are part of the record, which argues against treating the success as typical.
How defenders can read the report
Two details stand out for security teams. First, the operator's exposed Cairn console is the reason the campaign is documented at all, which suggests agent frameworks left open to the internet will keep leaking their operators' work. Second, the success depended on a service with authentication off by default, reachable over the network. Neither needs a new defensive idea, only old hygiene applied to tools that research groups and hospitals stand up quickly and forget.
Our take
The lesson is not that AI has made hacking effortless. It is that exposed tooling, default-off authentication and slow patch cycles now meet attackers who can try thousands of things for almost nothing. Watch for the MONAI Label fix, any CVE assignment from a second source, and whether Amazon or Google publish more detail on how often agents find flaws unaided.
Frequently asked questions
Can AI agents find zero-day vulnerabilities on their own?
Amazon Threat Intelligence reports one case: Cairn agents mapped MONAI Label's API and reached arbitrary file writes in about 11 minutes, and Amazon says it assesses with high confidence the flaw was a zero-day. It is one reported case, not a general rate.
What is Cairn?
Cairn is an open-source AI penetration testing framework that coordinates several agents. Amazon found an exposed, unauthenticated instance, which revealed the operator's prompts and agent output.
Who does Amazon blame for the attack?
An unidentified Chinese-nexus threat actor, with moderate confidence. BleepingComputer lists prompt language, timestamps and target choice as the basis. No operator is named.
Is MONAI Label patched?
Reports conflict. BleepingComputer says maintainers released a patch on November 18, 2026. SecurityWeek quotes advisory GHSA-qc22-rmgc-vg94 saying a fix will be in the next release. Check the maintainers' advisory directly.
What should MONAI Label users do?
SecurityWeek notes authentication is disabled by default and versions 0.8.5 and earlier are affected, so keep servers off untrusted networks, enable authentication and watch for the fixed release.
Sources
What each one is, and whose it is.
- 1
Chinese-nexus actor's Cairn AI agents found a zero-day on their own, AWS Security Blog (November 20, 2026)
Vendor announcement - 2
Amazon: China-linked hackers' AI agents exploited medical software zero-day, SecurityWeek (November 21, 2026)
Press reportIndependent of the vendor - 3
Hackers use autonomous AI tool Cairn to find zero-day, take over server, BleepingComputer (November 20, 2026)
Press reportIndependent of the vendor