What happened: Amazon says agents operating through Cairn found and exploited a previously unknown MONAI Label weakness during an attack on a university server. A later session used the same flaw to gain administrator-level control.
What matters now: As checked on November 21, the maintainers say a fix is in the code but list no patched release. They recommend restricting network access, enabling authentication and limiting permissions while administrators check their exposure.
Amazon Says Hackers’ AI Agents Found a Zero-Day in 11 Minutes
Amazon says attackers used AI agents to discover a previously unknown flaw in a medical-imaging application. The agents reached an initial breakthrough in about 11 minutes, but the more revealing problem is what happened after a fix was written.

Jump to a section
Eleven minutes was enough for an AI agent to find a way to write files where it should not have been allowed.
On November 20, Amazon Threat Intelligence described a September attack in which agents using an open-source framework called Cairn found a previously unknown weakness in a medical-imaging research application. Amazon links to the affected project’s advisory for MONAI Label, software used to help researchers label medical images. In a later session, the same weakness was used to gain root access, the highest level of control on the affected server.
The attacker was not absent. A person supplied targets and an objective, and later stepped in again. What changed was how much work the agents could perform between those instructions.
Amazon's account describes a zero-day vulnerability, a software weakness for which a fix was not available at the time of discovery. It does not establish that patient records were stolen, that hospitals using the application were compromised, or that this attack caused any of the recent service outages.
What happened in those 11 minutes
Amazon says it discovered the operation on September 10 after finding an unauthenticated Cairn server that exposed the operator's instructions and the agents' activity. The records showed a roughly 48-hour campaign, beginning September 8, against internet addresses associated with universities in East Asia outside mainland China.
Cairn coordinates AI agents, software systems that can use a model and tools to pursue an objective. In Amazon's account, the operator supplied target addresses and instructed the agents to gain shell access, a way to run commands on a computer. The agents then planned and adjusted their own steps within that task.
On one server, Amazon says, the agents identified the application, examined its interfaces and found a file-upload weakness. The flaw allowed a file to be written outside the directory where the application was supposed to store it. In about 11 minutes, the agents went from mapping that interface to making the unauthorized file write. That time does not measure the entire campaign or the time to full server control.
In a later session, the agents used the same weakness to place malicious code where the application would load it. Amazon says the result was a remote connection running as root. The distinction matters: the initial file-writing breakthrough and the later takeover were separate stages, not an 11-minute countdown to every possible consequence.
Amazon assessed with high confidence that the flaw was a zero-day at the time, because it found no corresponding fix, issue, advisory or prior vulnerability-database reference. The company assessed a China-linked actor with moderate confidence, based on factors including language, timestamps and targets. That is an attribution assessment, not proof of a particular person or government unit.
The fix exists. The release question is different.
The MONAI Label security advisory rates the flaw critical and lists versions through 0.8.5 as affected. It says the issue was fixed in a code change merged on October 19, but that the fix will be included in the next release.
As checked on November 21, the advisory lists no patched package version. The Python Package Index listing showed 0.8.5 as the latest stable release. A code change in a project's repository and an update installed on a production server are different things. Some installations may use a custom build that already contains the fix, so administrators need to verify what their own deployment actually runs.
BleepingComputer described a security update released on November 18. The advisory supports a narrower description: a fix is in the source code, while the listed package has no patched release yet. VentureBeat also identified this distinction. Release status can change, so check the maintainer's current advisory before acting.
That gap is the overlooked part of this story. A fix can exist in a public code repository while the ordinary package update still installs an affected version. Even after a release ships, someone must identify the exposed server, test the change where necessary and deploy it.
The dates show why those steps should not be treated as one event:
- September 8 to 10: Amazon says the observed attack campaign took place.
- October 19: The project merged the code change that its advisory identifies as the fix.
- November 18: The maintainer published the security advisory.
- November 20: Amazon published its account of the attacks.
- November 21: When Hacked.com checked, the advisory still listed no patched package version.
These dates do not establish when Amazon notified the maintainers or why a package had not been released. They show that a merged fix, a public advisory and an installable update are separate milestones.
What MONAI Label administrators should do
If your organization runs MONAI Label, involve the team responsible for the application and its infrastructure. Do not assume a server is safe merely because it is used for research, or compromised merely because it uses the software.
- Find the deployment and check its version. Identify servers, containers and externally managed instances. Compare the installed package or custom build with the current maintainer advisory and confirm whether the actual code includes the fix.
- Remove unnecessary exposure. The maintainers recommend keeping the server off untrusted networks. If the service is not needed, consider taking it offline through your normal operational process. Coordinate with the people who depend on it.
- Enable authentication and restrict access. The advisory says authentication is disabled by default. When it is enabled, the relevant annotator role can still reach the affected functionality, so do not grant that role to untrusted users.
- Limit what the service can change. Run it without administrator privileges and restrict write access to its intended data directory, as the maintainers recommend. These steps reduce exposure; they do not prove the server was never accessed.
- Check for earlier compromise. Ask the responsible security team or provider to review relevant logs, application changes and account activity. Preserve useful evidence, but do not delay urgent containment to collect it. If suspicious activity is found, use your incident-response process and involve qualified help.
- Track the release and update safely. When a supported release containing the fix is available, plan and verify the deployment. A custom build or package from an unverified source is not a substitute for a trusted update process.
For a broader plan covering exposed services, account access and recovery, see our business security guide and business incident response steps.
When AI finds the flaw before anyone publishes it
Our earlier report, The Patch Became a Map, examined Google's finding that an attack system could study published software fixes and use them to develop exploits. The Cairn case describes a different problem: Amazon says these agents found a weakness before there was a known fix.
The two cases are separate. Neither establishes that the same attackers, victims or software were involved. Together, they show why there are two clocks for defenders: the time before a weakness becomes known and the time between a fix becoming available and a system actually receiving it.
Google's Project Zero has said AI-assisted discovery and exploitation are among the reasons it is shortening its own vulnerability disclosure timelines. It plans to move its disclosure deadline to 60 days in February 2027 and 30 days in January 2028. Those are planned disclosure deadlines for Project Zero's reports, not a requirement that all software vendors meet those dates, and not a rule governing the MONAI case.
Our coverage of the FBI's ARTEX warning and stolen AI keys used in attacks describes other investigations. They help explain the wider use of agentic tools and computing resources, but they do not establish a connection to the Cairn operator.
The attacks that did not work
Amazon's report is not a story of an unstoppable machine. In one operation, agents spent about 16 hours taking 148 sequential actions against a server without succeeding. In another, they tried more than 30,000 username and password combinations over more than six hours and did not find a working login.
Those failures do not erase the successful intrusion. They do show that access controls, limited privileges and other barriers can make a difference, even when an attacker can keep trying. Amazon recommends high-confidence detections and automated responses where appropriate, with people overseeing the results. Automation can help defenders move faster, but a poorly chosen automatic action can also interrupt a legitimate service. Plan and test it against the systems you actually operate.
Most people do not need to check a medical-imaging server at home. The useful question for an organization is more concrete: which systems can be reached from outside, who owns them, and what happens when the fix exists but the update is not ready?