Thank you for reading or listening to The Realist Juggernaut. Independent journalism should be accessible to everyone.
SEATTLE, WASHINGTON — Amazon Threat Intelligence has identified a Chinese-linked cyber threat actor using an autonomous artificial intelligence framework known as Cairn to conduct cyberattacks against universities in East Asia. The investigation, disclosed November 20, 2026, revealed that the AI agents independently discovered an apparent zero-day vulnerability in an open-source medical imaging application and exploited it to obtain privileged access to a university server.
The findings were published by CJ Moses, Chief Information Security Officer of Amazon Integrated Security, following an investigation that began on September 10, 2026. Amazon researchers discovered a publicly accessible server operating Cairn, an open-source autonomous penetration-testing framework. The system had been deployed without authentication protections, allowing investigators to examine its operational history, including attack configurations, instructions, execution records, and responses generated by the AI agents.
The exposed records documented a campaign that began around September 8 and continued for approximately 48 hours. Targeted internet addresses belonged to universities located in East Asia outside mainland China. The operator submitted instructions in Simplified Chinese, while scripts used during the campaign contained timestamps configured to China Standard Time. Amazon assessed with moderate confidence that the activity originated from a Chinese-linked threat actor, although its findings did not establish that the campaign was directed by the Chinese government.
Cairn was originally developed as a research and penetration-testing framework capable of coordinating multiple AI agents to accomplish technical objectives. Unlike conventional automation that follows a predetermined sequence of commands, the framework can interpret natural-language instructions, evaluate the results of its actions, and determine additional steps without requiring an operator to manually direct every operation.
During the investigated campaign, the attacker supplied targeted IP addresses and instructed the system to obtain shell access. The agents then performed reconnaissance, examined exposed services, evaluated potential weaknesses, and attempted to establish control over the selected environments.
One of the most extensive operations targeted an exposed Redis database service at a university. Over approximately 16 hours, the agents executed 148 sequential actions in attempts to compromise the system. Their activity included testing Redis replication behavior, attempting to escape scripting restrictions, manipulating configuration files, and examining weaknesses that could permit unauthorized code execution.
The agents also attempted to exploit the Log4Shell vulnerability against another service and investigated a separate Qdrant instance through a server-side request forgery weakness. That activity enabled exploration of the internal Docker network environment, followed by attempts against PostgreSQL, Secure Shell services, and other network components. Amazon reported that these efforts failed to achieve the intended compromise.
A second operation targeted a university server equipped with graphics processing units used for computationally intensive research. The agents gathered approximately 80 legitimate usernames from publicly available faculty directories, laboratory websites, archived institutional pages, and academic publications. They then tested more than 30,000 credential combinations against remote-access services and other university infrastructure during an operation lasting over six hours.
The credential attempts extended beyond the primary server to administrative interfaces, network storage, virtual private network infrastructure, and the institution’s campus portal. None of the tested combinations produced valid access. The operation demonstrated the ability of autonomous agents to sustain high-volume reconnaissance and authentication attempts while continuously processing information collected from the target environment.
The most consequential finding emerged when Cairn encountered a university system running an open-source medical imaging AI application. The agents identified the application’s underlying framework, mapped its available application programming interface endpoints, and examined a file-upload mechanism that accepted user-controlled filenames.
Within approximately 11 minutes, the agents identified a path traversal vulnerability that allowed uploaded files to be written outside their intended storage directory. This class of vulnerability occurs when an application fails to restrict file paths to authorized locations, potentially allowing an attacker to manipulate files elsewhere on the server.
In a later session, the agents used the weakness to place malicious Python code in a location automatically loaded when a Python interpreter starts. By triggering an application function that launched a new Python process, the agents caused the malicious code to execute and established a reverse shell with root-level privileges.
That result gave the attacker-controlled process extensive authority over the compromised server. Root access can permit modification of system files, execution of additional programs, interference with services, and access to information available to the privileged account. The actual consequences depend upon the configuration and contents of the affected environment.
Amazon determined with high confidence that the vulnerability qualified as a zero-day at the time of discovery. Investigators found no corresponding public security advisory, documented correction, or previously reported vulnerability record for the affected open-source project. The finding is significant because the agents were not relying exclusively on a known vulnerability catalog; they identified a previously undocumented weakness through automated examination of the application.
The affected software was identified as MONAI Label, an open-source medical imaging annotation and artificial intelligence application. Its maintainers published security advisory GHSA-qc22-rmgc-vg94, documenting an arbitrary file-write vulnerability caused by insufficient validation of file paths within datastore upload operations.
The advisory identifies the issue as CWE-22, a path traversal weakness, and assigns it a critical severity classification. Versions through 0.8.5 are listed as affected. Although the underlying code has been corrected through a repository change, the advisory does not yet identify a patched package release and states that the correction will be included in a future MONAI Label version.
The security implications are amplified by the application’s authentication configuration. According to the maintainers, authentication is disabled by default, meaning that a network-accessible deployment using the affected configuration could permit an unauthorized user to exploit the file-handling weakness. Where authentication is enabled, users assigned the ANNOTATOR role may still possess sufficient permissions to reach the vulnerable functionality.
Amazon also identified activity involving a second university network where the same vulnerable application was present. Following the initial discovery, the human operator supplied a proxy connection into that environment and directed the agents to investigate additional systems. Amazon inferred that the vulnerability had already been exploited to establish the proxy connection, although the published evidence does not independently establish the complete sequence of the earlier compromise.
Amazon notified the MONAI Label maintainers of the vulnerability. Because the affected universities were not AWS customers and Amazon lacked a direct notification channel, the company coordinated with a government partner to alert the organizations. The universities and their countries were not publicly identified.
The investigation follows additional reporting from Amazon concerning another autonomous offensive framework, ARTEX, which investigators observed being used in credential theft and other cybercriminal operations.
Google Threat Intelligence Group has separately documented malicious deployment and experimentation involving Cairn, ARTEX, PentAGI, and CyberStrikeAI. Google’s findings include Cairn activity involving targets in the United States aerospace sector, demonstrating that the framework’s observed use extends beyond the university campaign identified by Amazon.
For organizations operating MONAI Label, the published security advisory recommends restricting access from untrusted networks, enabling authentication, limiting permissions granted to application users, and operating the software with only the filesystem privileges necessary for its intended functions. Administrators should also monitor the project’s security releases to obtain the corrected package when available.
Amazon’s findings reinforce the importance of continuous security monitoring and automated defensive response. Traditional incident-response processes built around manual alert examination can face operational pressure when an attacker executes reconnaissance and exploitation within minutes. Effective defenses must be capable of identifying suspicious behavior, restricting unauthorized activity, and preserving forensic evidence without depending exclusively on human intervention at every stage.
Official Amazon investigation documenting a Chinese-linked threat actor’s use of autonomous Cairn AI agents to target universities in East Asia. The report details prolonged intrusion attempts, the discovery and exploitation of a previously unknown vulnerability, unauthorized root-level server access, and Amazon’s coordination with software maintainers and government partners. (Free Download)
Official Google Threat Intelligence Group report examining the use of autonomous AI agents and offensive cybersecurity frameworks, including Cairn, ARTEX, CyberStrikeAI, and PentAGI. The findings document AI-assisted vulnerability exploitation, cyber espionage operations, automated attack infrastructure, and activity involving targets in the United States aerospace sector. (Free Download)
🔥 NOW AVAILABLE! 🔥
🔥 NOW AVAILABLE! 🔥
📖 INK & FIRE: BOOK 1 📖
A bold and unapologetic collection of poetry that ignites the soul. Ink & Fire dives deep into raw emotions, truth, and the human experience—unfiltered and untamed
🔥 Kindle Edition 👉 https://a.co/d/9EoGKzh
🔥 Paperback 👉 https://a.co/d/9EoGKzh
🔥 Hardcover Edition 👉 https://a.co/d/0ITmDIB
🔥 NOW AVAILABLE! 🔥
📖 INK & FIRE: BOOK 2 📖
A bold and unapologetic collection of poetry that ignites the soul. Ink & Fire dives deep into raw emotions, truth, and the human experience—unfiltered and untamed just like the first one.
🔥 Kindle Edition 👉 https://a.co/d/1xlx7J2
🔥 Paperback 👉 https://a.co/d/a7vFHN6
🔥 Hardcover Edition 👉 https://a.co/d/efhu1ON
Get your copy today and experience poetry like never before. #InkAndFire #PoetryUnleashed #FuelTheFire
🚨 NOW AVAILABLE! 🚨
📖 THE INEVITABLE: THE DAWN OF A NEW ERA 📖
A powerful, eye-opening read that challenges the status quo and explores the future unfolding before us. Dive into a journey of truth, change, and the forces shaping our world.
🔥 Kindle Edition 👉 https://a.co/d/0FzX6MH
🔥 Paperback 👉 https://a.co/d/2IsxLof
🔥 Hardcover Edition 👉 https://a.co/d/bz01raP
Get your copy today and be part of the new era. #TheInevitable #TruthUnveiled #NewEra
🚀 NOW AVAILABLE! 🚀
📖 THE FORGOTTEN OUTPOST 📖
The Cold War Moon Base They Swore Never Existed
What if the moon landing was just the cover story?
Dive into the boldest investigation The Realist Juggernaut has ever published—featuring declassified files, ghost missions, whistleblower testimony, and black-budget secrets buried in lunar dust.
🔥 Kindle Edition 👉 https://a.co/d/2Mu03Iu
🛸 Paperback Coming Soon
Discover the base they never wanted you to find. TheForgottenOutpost #RealistJuggernaut #MoonBaseTruth #ColdWarSecrets #Declassified



