The stopwatch is the story
Amazon’s new incident report is alarming less because attackers used AI and more because of the speed Amazon says it observed. In a campaign against universities in East Asia outside mainland China, autonomous Cairn agents were given a simple instruction: “obtain shell access.” They then planned, tested and adapted without the operator choosing each next step, Amazon said .
The successful branch of the operation targeted MONAI Label, an open-source medical-imaging AI application used to assist researchers with annotation workflows. Amazon says the agents identified the application framework, mapped API endpoints, found an upload function with a filename parameter, tested it for path traversal and moved from API mapping to arbitrary file writing in about 11 minutes .
That “about 11 minutes” is the hard number defenders will remember. It compresses the familiar sequence of vulnerability discovery, exploit development and intrusion into a window shorter than many security operations centers need to open, triage and assign a ticket. Amazon’s CISO CJ Moses framed the lesson bluntly: an agent that can discover and exploit a zero-day in minutes can finish its activity faster than many organizations can manually analyze an alert .
What Amazon says happened
Amazon Threat Intelligence said it discovered the operation on September 10, 2026, when it found a publicly reachable Cairn server running without authentication. Because the attacker’s own platform was exposed, Amazon could view operational history, including project configurations, prompts and AI outputs .
The campaign had begun roughly 48 hours earlier, on September 8, and targeted IP addresses associated with universities in East Asia outside mainland China . Amazon attributed the activity with moderate confidence to a Chinese-nexus threat actor, citing Simplified Chinese prompts, script timestamps set to China Standard Time and the selection of university targets .
Cairn is described by Amazon and security reporters as an open-source autonomous penetration-testing framework that coordinates multiple AI agents toward a plain-language goal . In legitimate hands, that kind of framework can help defenders test systems. In this case, Amazon says the operator supplied targets and a goal, then let the agents work through tactics.
Not every attempt succeeded. One operation against a Redis server ran for about 16 hours and 148 sequential actions, including attempts involving Redis features, Log4Shell and follow-on internal mapping; it failed . Another operation against a university GPU server ran for more than six hours, built a list of about 80 real usernames from public sources and tested more than 30,000 credential combinations; that failed too .
The MONAI Label case was different. Amazon says the agents found a previously unknown path traversal vulnerability and used it to write files outside the intended application area . In a later session, they used the same weakness to establish a Python reverse shell running as root, giving the attacker full control of the server .
The vulnerable tool: MONAI Label
MONAI Label matters because it sits in a sensitive workflow. It is not a generic demo app; it supports medical-imaging annotation, the kind of research environment where universities, hospitals and AI labs often share tools, datasets and compute resources. BleepingComputer identified the affected product as MONAI Label and described it as a medical-imaging application that uses AI to help researchers label medical images .
The project’s GitHub Security Advisory, published on November 18, 2026, names the issue as a path traversal flaw in datastore upload endpoints that can allow arbitrary file write . The advisory rates the issue critical, says the affected package is monailabel, and lists affected versions as 0.8.5 and earlier .
The advisory says no patched version is currently listed, even though the fix has been made in the project code and is expected in the next MONAI Label release . VentureBeat separately reported that MONAI Label 0.8.5 remained the newest PyPI release and that no patched release was available at publication time .
That distinction matters. “Fixed in code” is not the same as “available to install.” For operators of exposed MONAI Label servers, the current state is a mitigation problem, not a routine upgrade problem.
Why Amazon calls it a zero-day
Amazon says it assesses with high confidence that the MONAI Label flaw was a zero-day because, at the time of discovery, it found no corresponding fix, issue, advisory or prior mention in vulnerability databases for the affected open-source project . The GitHub advisory also states that there is no known CVE for the issue .
The term “zero-day” is sometimes overused, but here the timeline is central. The agents were not merely applying a public exploit after a disclosure. According to Amazon, they found an unpatched vulnerability during an active intrusion attempt, weaponized it and used it before a public advisory and before an installable patched release was available .
The advisory describes why the bug is dangerous without requiring a named CVE: a caller-controlled image identifier or label path could be used to write uploaded content outside the intended datastore, and arbitrary file write can lead to remote code execution or denial of service depending on where the MONAI Label process is allowed to write .
The uncomfortable role of defaults
The GitHub advisory says authentication is disabled by default, meaning that in the default configuration any network client able to reach the server can exploit the flaw . Even when authentication is enabled, the advisory says the ANNOTATOR role is sufficient to exploit it .
This turns an application-level bug into an exposure-management problem. A MONAI Label instance on a private research network is one risk. A MONAI Label instance reachable from untrusted networks, running with broad write permissions, is another. Amazon’s observed case also shows why running research tools with excessive privileges can convert a file-write vulnerability into full system compromise .
Until an upgradeable release is available, the advisory recommends enabling authentication, avoiding the ANNOTATOR role for untrusted users, keeping MONAI Label off untrusted networks and running the service as an unprivileged user with write access limited to the datastore directory .
The second network and what is still unknown
Amazon says the human operator later stepped in after the agents discovered the zero-day, providing a proxy that tunneled into a second university network. The proxy connected to a server running the same vulnerable software, and Amazon inferred that the actor had used the zero-day to compromise it and establish the tunnel .
The public reports do not name the universities, their countries or the data exposed, if any. The Record reported that AWS declined to identify the universities or country involved and said Amazon worked with a government partner to notify affected organizations [5]. That leaves important questions open: whether protected research data was accessed, whether medical images were touched, whether credentials were harvested and whether the second environment suffered broader lateral movement.
Those unknowns should not soften the operational lesson. The disclosed evidence is already enough to show that autonomous offensive tooling can operate at a tempo that breaks the comfortable assumption that defenders will get a patch, test it and deploy it before exploitation arrives.
What defenders should do now
For teams running MONAI Label, the priority is immediate exposure review. Identify deployments of version 0.8.5 and earlier, check whether the service is reachable from untrusted networks, verify authentication settings and review whether the process runs with unnecessary privileges .
Administrators should also assume that simply enabling authentication today does not rule out earlier compromise. Because Amazon says the flaw was used in the wild, teams should review host and application logs for unexpected file writes, abnormal Python process launches, new scheduled tasks, suspicious SSH authorization changes and outbound callbacks from MONAI Label hosts .
For security leaders, the broader response is architectural. Manual triage is still necessary, but it cannot be the first and only control when adversary agents can run continuously, fail for 16 hours, pivot and then succeed in 11 minutes on a different target . High-confidence detections need automated containment paths: isolate the host, revoke exposed credentials, block egress, snapshot evidence and escalate to humans with context rather than raw alert noise.
The blue-team joke writes itself: Skynet does not wait for Patch Tuesday. But the practical takeaway is more sober. The old clock measured days between disclosure and exploitation. Amazon says this clock measured minutes .


Comments
Be the first to comment.