# Autonomous AI Agents Exploit Medical Imaging Platform via Open-Source Cairn Framework

> Source: <https://epyhia.com/autonomous-ai-agents-exploit-medical-imaging-platform-via-open-source>
> Published: 2026-11-22 12:09:33+00:00

**Recommendation:** Treat unauthenticated open-source services like MONAI Label and Redis as agent-reachable attack surface: patch path-traversal upload parameters and require authentication, since Cairn agents probed 148 actions and 30,000+ credentials without human direction.
More in the playbook "Securing agents and AI tools": <https://wpnews.pro/playbooks/agent-security>

# Autonomous AI Agents Exploit Medical Imaging Platform via Open-Source Cairn Framework

Amazon Threat Intelligence has documented one of the first publicly detailed cases of autonomous AI agents discovering and exploiting a zero-day vulnerability during a live intrusion campaign. The agents, running the open-source Cairn penetration-testing framework, achieved arbitrary file write on a university server running MONAI Label, an open-source medical imaging annotation platform, in approximately 11 minutes. A root shell was obtained in a subsequent session. Amazon attributes the operation to a Chinese-nexus threat actor with moderate confidence.

## Discovery and campaign overview

On September 10, 2026, Amazon researchers found a publicly accessible, unauthenticated instance of Cairn — an open-source framework that coordinates multiple AI agents toward a goal defined in plain language. The instance contained the operator's prompts, written in Simplified Chinese, and script timestamps set to China Standard Time. The operational history revealed a roughly 48-hour campaign beginning September 8 against IP addresses associated with universities in East Asia outside mainland China. The operator's stated objective was simply "obtain shell access" [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own).

Cairn was originally built as a research and problem-solving engine for defenders. Given a starting point and a goal, it plans, executes, and adjusts its approach without the operator choosing each step. That same ease of use makes it attractive to threat actors. Gambit Security has separately documented a financially motivated group using Cairn against online retailers, indicating the framework is being adopted in both espionage and criminal operations [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own).

## Failed operations show persistence

Not every operation succeeded. Against one university's unauthenticated Redis instance, the agents spent approximately 16 hours and 148 sequential actions attempting known techniques: replication abuse, scripting sandbox escape, configuration file writes, memory corruption, and Log4Shell injection against a co-located service. When those failed, the agents mapped the internal Docker network via a server-side request forgery vulnerability in a Qdrant instance and launched credential sprays against PostgreSQL, SSH, and other services. None succeeded [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own).

A separate operation ran for more than six hours against a university GPU server. The agents first compiled roughly 80 real usernames from faculty directories, lab websites, archived pages, and academic papers, then tested more than 30,000 credential combinations against SSH, an admin dashboard, network storage, a VPN server, and the campus portal. They never found a working login [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own).

The AWS blog noted that the agents "kept adapting for nearly a full day without human direction" — a pace and persistence that changes the economics of intrusion when patient, iterative probing requires little human labor [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own). Amazon CISO CJ Moses separately observed that the framework "is being adopted in both espionage and criminal operations" [\[7\]](https://labs.cloudsecurityalliance.org/research/csa-research-note-monai-label-agent-discovered-zero-day-2026).

## Zero-day discovery and exploitation

Against a target running MONAI Label, the outcome was different. The agents identified the application framework, mapped its API endpoints, and found an upload endpoint with a filename parameter vulnerable to path traversal. In about 11 minutes, they went from mapping the application's API to arbitrarily writing files on the server [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own).

In a later session, the agents used the same vulnerability to write a Python reverse shell payload to `sitecustomize.py`, a file Python automatically imports at interpreter startup. They then triggered a training API endpoint to spawn a new Python process, which imported the payload and established a reverse shell running as root [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own).

Amazon assesses with high confidence that this was a zero-day at the time of exploitation, based on the absence of any corresponding fix, issue, advisory, or prior mention in vulnerability databases for the affected open-source project [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own). The vulnerability is tracked as GitHub advisory GHSA-qc22-rmgc-vg94 and affects MONAI Label versions 0.8.5 and earlier [\[8\]](https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94). A third-party analysis by Aviatrix refers to it as CVE-2026-133968 (CVSS 9.8) [\[4\]](https://aviatrix.ai/threat-research-center/attackers-ai-agents-zero-day-medical-imaging-cve-2026-133968), but as of November 21 the Cloud Security Alliance reported that no CVE identifier had been assigned [\[7\]](https://labs.cloudsecurityalliance.org/research/csa-research-note-monai-label-agent-discovered-zero-day-2026), and the GitHub advisory shows no CVE ID [\[8\]](https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94).

## The vulnerability and patch gap

The flaw resides in the `PUT /datastore/image` and `PUT /datastore/label` endpoints. The `image` query parameter is used as an identifier without validation; `LocalDatastore.add_image()` appends a file extension, joins it to the datastore directory, and writes the uploaded content via `shutil.copy()` without checking that the resolved path stays inside the datastore. A value such as `../../../tmp/x` writes outside the intended storage directory. Authentication is disabled by default (`MONAI_LABEL_AUTH_ENABLE=False`), so any network client that can reach the server can exploit it. With authentication enabled, the low-privilege `ANNOTATOR` role is still sufficient [\[8\]](https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94).

A fix was merged in pull request #1915 on October 19, 2026, which resolves destination paths and verifies they remain inside the datastore directories, while the API endpoint now strips directory components from the image ID and rejects empty or dot-prefixed values with HTTP 400 [\[10\]](https://github.com/Project-MONAI/MONAILabel/pull/1915). However, as of November 20, no patched release exists on PyPI; version 0.8.5, published in November 2024, remains the newest installable package. No CVE identifier had been assigned at the time of the advisory publication on November 18 [\[7\]](https://labs.cloudsecurityalliance.org/research/csa-research-note-monai-label-agent-discovered-zero-day-2026)[\[8\]](https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94).

This gap — 33 days and counting between a public fix commit and an installable package — leaves defenders dependent on compensating controls. The MONAI Label advisory (GHSA-qc22-rmgc-vg94) lists workarounds: enable authentication, restrict the `ANNOTATOR` role, do not expose the server to untrusted networks, and run the service as an unprivileged user with write access limited to the datastore directory [\[8\]](https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94).

## Subsequent pivot and attribution

After the agents discovered the zero-day, the operator stepped in, providing a proxy that tunneled into a second university network. The server at the other end of the tunnel ran the same vulnerable software, and Amazon infers the actor had used the zero-day to compromise it and establish the tunnel. The operator then directed the agents to target other machines in that network [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own).

Amazon attributes the activity to a Chinese-nexus threat actor with moderate confidence, citing the Simplified Chinese prompts, China Standard Time timestamps, and the targeting of universities likely of strategic interest to China [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own). The company notified the MONAI Label maintainers and worked with a government partner to notify the affected organizations, which were not AWS customers [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own).

## What this means for defenders

The operational tempo compresses the window for defensive response. An agent that discovers and exploits a zero-day in minutes completes its activity faster than many security operations centers can manually triage an alert. The agents that failed against the hardened Redis target ran continuously for 16 hours without breaks, sleep, or working-hour constraints [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own).

The Cloud Security Alliance notes that research-grade medical AI tooling is frequently deployed by small academic teams with default settings, often missing from asset inventories and vulnerability scanning, and may run on GPU hosts with elevated privileges and broad network reach into imaging archives [\[7\]](https://labs.cloudsecurityalliance.org/research/csa-research-note-monai-label-agent-discovered-zero-day-2026). The same agents probed Redis, Qdrant, and PostgreSQL — services commonly found in AI research environments — not only AI-specific tools [\[7\]](https://labs.cloudsecurityalliance.org/research/csa-research-note-monai-label-agent-discovered-zero-day-2026).

Organizations running MONAI Label should treat every reachable instance as potentially exposed, apply the advisory's workarounds immediately, and review logs for traversal patterns against the datastore endpoints and the source addresses Amazon published (142.4.98.145 and 142.4.107.34, active September 8–10, 2026) [\[1\]](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own)[\[7\]](https://labs.cloudsecurityalliance.org/research/csa-research-note-monai-label-agent-discovered-zero-day-2026). Because the flaw is exploitable without credentials in default deployments and has been publicly documented, any instance that was reachable before mitigation should be treated as potentially compromised [\[7\]](https://labs.cloudsecurityalliance.org/research/csa-research-note-monai-label-agent-discovered-zero-day-2026).

## Sources

1. [Chinese-nexus actor's Cairn AI agents found a zero-day on their own](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own) (aws.amazon.com)
2. [China-Linked Hackers' AI Agents Exploited Medical Software Zero-Day](https://www.securityweek.com/amazon-china-linked-hackers-ai-agents-exploited-medical-software-zero-day) (securityweek.com)
3. [Hackers Use Cairn to Hack Medical Imaging Application](https://www.govinfosecurity.com/hackers-use-cairn-to-hack-medical-imaging-application-a-33469) (govinfosecurity.com)
4. [CVE-2026-133968: AI Agents Exploit MONAI Label Zero-Day for Root ...](https://aviatrix.ai/threat-research-center/attackers-ai-agents-zero-day-medical-imaging-cve-2026-133968) (aviatrix.ai)
5. [How an Open-Source AI Agent Found and Exploited a Zero-Day (And Got ...](https://medium.com/@kankit570/how-an-open-source-ai-agent-found-and-exploited-a-zero-day-and-got-caught-by-a-missing-password-bfcc6c4d6d4f) (medium.com)
6. [Cairn Agents Exploit MONAI Label Zero-Day at East Asian Universities](https://brandefense.io/security-news/cairn-monai-label-zero-day-universities) (brandefense.io)
7. [MONAI Label Zero-Day: AI Agents Exploit Medical Imaging Research Servers – Lab Space](https://labs.cloudsecurityalliance.org/research/csa-research-note-monai-label-agent-discovered-zero-day-2026) (labs.cloudsecurityalliance.org)
8. [Path traversal in datastore upload endpoints allows arbitrary file write · Advisory · Project-MONAI/MONAILabel · GitHub](https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94) (github.com)
9. [Your Privacy Choices Page 1](https://pulse.aws/application/ZRPLWLL6?p=0) (pulse.aws)
10. [fix: confine datastore write paths by garciadias · Pull Request #1915 · Project-MONAI/MONAILabel](https://github.com/Project-MONAI/MONAILabel/pull/1915) (github.com)
