AI Agents Found a Zero-Day in 11 Minutes. The Real Risk Is What Came Next

A suspected Chinese-nexus actor used autonomous Cairn AI agents to discover and exploit a critical MONAI Label vulnerability during attacks against universities, showing how agentic tools can compress parts of an intrusion into minutes.

QuantumBytz Team
November 21, 2026
Share:
Autonomous AI agents mapping and compromising a medical-imaging application server during a zero-day attack.

Summary

A suspected Chinese-nexus actor used autonomous Cairn AI agents to find an unknown vulnerability in MONAI Label, gain arbitrary file-write access and eventually take control of a university server. The incident shows how AI changes the speed and persistence of offensive operations, even when many individual attack attempts fail.

Autonomous AI agents have crossed an important line in offensive cybersecurity.

According to Amazon Threat Intelligence, a suspected Chinese-nexus actor used the open-source Cairn framework to attack university systems with minimal human direction. The operator supplied target addresses and a simple objective: “obtain shell access.”

The agents handled much of the work themselves. They mapped services, tested possible weaknesses, changed tactics when attempts failed and continued operating for hours without waiting for a person to select every command.

Against one target, they found a previously unknown path-traversal vulnerability in MONAI Label, an open-source application used for AI-assisted medical-image annotation. Amazon says the agents progressed from mapping the application’s API to writing files outside the intended storage directory in about 11 minutes.[1]

In a later session, the attackers used that access to plant a reverse shell and obtain root-level control of the server.[1]

This is not a theoretical demonstration or a laboratory benchmark. The vulnerability was found during a real intrusion campaign.

How Amazon uncovered the operation

Amazon discovered the campaign on September 10 after finding a publicly accessible Cairn server that had been left without authentication. Because the operator failed to secure it, Amazon could inspect the framework’s operational history, including project configurations, prompts and agent output.[1]

The records showed an approximately 48-hour campaign against universities in East Asia outside mainland China. Prompts were written in Simplified Chinese, while scripts used China Standard Time.

Amazon attributes the operation to a Chinese-nexus actor with moderate confidence. That assessment is based on the language, timestamps and selection of targets, rather than proof tying the activity to a specific government agency.[1]

The exposed history also provided an unusually clear view of what autonomous attack agents can and cannot currently do.

They were persistent, but they were not magically successful against everything.

One operation ran for roughly 16 hours and involved 148 sequential actions against a Redis instance. The agents tried several known takeover methods, attempted Log4Shell against another service and mapped part of the internal container network. None of those attempts produced a successful compromise.[1]

Another effort spent more than six hours testing tens of thousands of credential combinations associated with real usernames. That operation also failed.[3]

The successful MONAI Label attack stands out because the agents did not merely automate a known public exploit. They identified the application, mapped its interfaces, found a vulnerable parameter and tested it for path traversal.

The MONAI Label vulnerability

The affected code is in MONAI Label’s datastore upload functionality.

According to the project’s security advisory, the application accepted an image identifier supplied by the client and used it to construct a filesystem path. It did not verify that the resolved destination remained inside the intended datastore directory.[2]

An attacker could therefore supply a manipulated path and write an uploaded file somewhere else on the server, provided the MONAI Label process had permission to write there.

Arbitrary file writes are dangerous because the attacker may be able to overwrite application code, alter configuration files, establish persistence or place content where another process will execute it. In this campaign, the attackers eventually turned the file-write primitive into root access.[1][2]

The advisory classifies the vulnerability as critical and lists MONAI Label 0.8.5 and earlier as affected. It currently lists no patched release. The correction has been merged into the project’s source code and is expected to appear in the next MONAI Label release.[2]

The default deployment makes the situation worse: authentication is disabled unless an administrator explicitly enables it. When authentication is enabled, accounts assigned the ANNOTATOR role may still reach the vulnerable functionality.[2]

At the time the agents found and exploited the weakness, Amazon says there was no corresponding advisory, fix, public issue or entry in vulnerability databases. That is why Amazon assesses with high confidence that it was a zero-day.[1]

It is no longer an undisclosed vulnerability, but organizations running the currently released package may still be exposed.

Why the 11-minute figure matters

The important number is not simply that the agents needed about 11 minutes.

A skilled human researcher could also identify a path-traversal bug quickly under favorable conditions. The larger change is that the agents could perform reconnaissance, test hypotheses and continue operating without the normal human pauses between each stage.

They did not need sleep. They did not become bored after several hours of failed credential attempts. They could work through one approach, abandon it and immediately try another.

That changes the economics of broad offensive campaigns.

A threat actor no longer needs a large team to give sustained attention to every target. Autonomous frameworks can perform the repetitive work across many systems, then concentrate effort where they find an opening.

The same tools will still make mistakes. Amazon’s evidence includes long, unsuccessful operations. But failure becomes cheaper when software can keep trying without consuming an operator’s full attention.

Defenders should expect the interval between exposure, vulnerability discovery and exploitation to continue shrinking.

This is an exposure-management problem too

The attack was not powered by AI alone.

Several ordinary security failures made it possible:

  • The attacker’s Cairn control server was itself exposed without authentication.
  • Target services were reachable over the network.
  • MONAI Label disables authentication by default.
  • The application process had enough filesystem permission for the flaw to become dangerous.
  • The eventual malicious code executed with root privileges.

AI accelerated the discovery and exploitation process, but familiar infrastructure weaknesses still determined the outcome.

That distinction matters. Organizations do not need an “anti-AI firewall” to address the immediate risk. They need reliable inventories, restricted network exposure, strong authentication and tightly scoped service permissions.

An application handling medical data or GPU workloads should not automatically run as root. An annotation role should not carry the practical ability to write executable content outside the application’s datastore. An internal research service should not be reachable from untrusted networks merely because deployment was easier that way.

What MONAI Label operators should do

MONAI Label administrators should first determine whether any instance running version 0.8.5 or earlier is reachable from the internet, an untrusted research network or a broadly accessible VPN segment.

Until a fixed release is available, the project recommends several mitigations:[2]

  1. Enable authentication with MONAI_LABEL_AUTH_ENABLE=True.
  2. Do not assign the ANNOTATOR role to untrusted users.
  3. Remove the server from untrusted networks.
  4. Run the application under an unprivileged account.
  5. Limit write access to the required datastore directories.

Applying the source-level fix or installing a future patched release addresses the vulnerable path handling, but organizations should not stop there.

Because this flaw was exploited before public disclosure, exposed systems should be treated as potentially compromised. Administrators should inspect them for unexplained file changes, unexpected startup hooks, new scheduled tasks, modified SSH authorization files and unfamiliar outbound connections.

If the application ran as root, the investigation should extend beyond the MONAI Label directory. Reinstalling the package does not remove persistence elsewhere on the host.

What security teams should learn from the campaign

The incident changes the assumptions behind detection and response.

Traditional monitoring often expects a human-paced sequence: reconnaissance, initial access, persistence and lateral movement over hours or days. Autonomous agents can compress parts of that sequence into minutes while simultaneously generating a large amount of noisy or unsuccessful activity.

A security operations center may still be investigating the first alert when the attack framework has already changed direction.

That makes high-confidence automated containment more important. Examples include isolating a workload after suspicious arbitrary file writes, blocking unexpected child processes from an application service, revoking exposed credentials and preserving volatile evidence before the attacker can establish deeper persistence.

Human analysts remain necessary, but they cannot manually approve every defensive step if the attacker’s agents are operating at machine speed.

Security teams also need to watch the tools surrounding AI systems. An exposed agent framework may reveal prompts, credentials, targets and operational history, just as Amazon observed here. Offensive and defensive agents should be treated as privileged infrastructure, with authentication, network restrictions, audit logging and secrets management appropriate to their capabilities.

AI did not eliminate the need for expertise

It would be easy to describe this incident as proof that autonomous agents can replace experienced attackers. Amazon’s evidence does not support such a simple conclusion.

Most of the recorded operations failed. The operator still selected targets and defined the objective. Turning one discovery into a useful campaign still required decisions about access, persistence and where to move next.

What the agents changed was the amount of work one operator could sustain.

They could attempt many techniques, maintain pressure for hours and exploit a newly discovered weakness without requiring constant manual direction. That is enough to matter.

The near-term threat is not an all-knowing autonomous hacker. It is an ordinary threat actor gaining the persistence and parallelism of a much larger team.

The takeaway

The MONAI Label compromise is an early example of a security problem organizations will see more often: agents finding vulnerabilities during live operations rather than merely replaying known exploits.

The defensive answer is not panic about AI. It is reducing the number of easy opportunities those agents can test.

Services should not be publicly reachable by accident. Authentication should not be disabled by default on sensitive systems. Application processes should not have broad filesystem access or root privileges. Security teams should be able to contain high-confidence malicious behavior faster than a human analyst can manually investigate it.

Autonomous offensive tools will continue to improve. Infrastructure teams need to make sure that faster attackers keep finding fewer useful openings.

Sources

[1] AWS Security Blog — Cairn AI agents found a zero-day

[2] MONAI Label security advisory

[3] The Record — AI agents discovered zero-day during university attacks

[4] SecurityWeek — AI agents exploited medical software zero-day

QuantumBytz Team

The QuantumBytz Editorial Team covers cutting-edge computing infrastructure, including quantum computing, AI systems, Linux performance, HPC, and enterprise tooling. Our mission is to provide accurate, in-depth technical content for infrastructure professionals.

Learn more about our editorial team