Suspected Chinese state-linked actors have successfully deployed autonomous AI agents to identify and exploit a previously unknown vulnerability in an open-source medical imaging application. According to findings from Amazon’s Threat Intelligence Team, the campaign targeted university networks in East Asia throughout a 48-hour period beginning on September 8, 2026. The attackers utilized a penetration-testing framework known as Cairn, which allows AI agents to execute multi-stage cyberattacks without direct human intervention.
The operation was uncovered on September 10, when security researchers identified a server hosting an instance of the Cairn framework. Analysis of the activity logs revealed that the AI agents were configured with timestamps set to China Standard Time and utilized prompts written in Simplified Chinese. The agents demonstrated a high degree of persistence, adapting their tactics over the course of nearly a full day as they cycled through various failed intrusion attempts before successfully compromising a target.
Before discovering the zero-day vulnerability, the agents attempted a series of unsuccessful attacks against university infrastructure. These included a 16-hour effort to breach a Redis server, which involved 148 sequential actions such as memory corruption, configuration file manipulation, and attempts to escape the scripting sandbox. When these efforts failed, the agents pivoted to other techniques, including Log4Shell injection, server-side request forgery against a Qdrant instance, and credential spraying against SSH and PostgreSQL services. A separate, concurrent operation targeted a university GPU server for over six hours, utilizing a list of 80 legitimate usernames scraped from academic papers and public websites.
The breakthrough occurred when the agents identified a path traversal vulnerability within MONAI Label, an open-source framework used for medical imaging AI. Within approximately 11 minutes of mapping the application’s API endpoints, the agents successfully located an upload endpoint with a vulnerable filename parameter. This allowed them to write arbitrary files to the server and eventually execute a Python reverse shell with root-level privileges. Once the zero-day was exploited, the human operators intervened to provide a proxy, enabling the agents to tunnel into a second university network and compromise a server running the same vulnerable software.
The maintainers of MONAI Label have since released a security advisory and a patch to address the path traversal flaw following notification from Amazon. This incident is not an isolated case; it aligns with broader intelligence regarding the weaponization of AI frameworks. Separate research from Gambit Security has documented a Chinese operator utilizing the Cairn framework to conduct at least 105 attacks, resulting in the compromise of 27 companies.
These campaigns highlight a shift toward the use of autonomous AI harnesses for both espionage and financial crime. Beyond the university attacks, evidence suggests these frameworks have been deployed to target major commercial entities, including a Fortune 500 hospitality firm, a large industrial supplies distributor, a major airline, and a prominent online fashion retailer. In those instances, the agents were reportedly used to exfiltrate sensitive data, including credit card records.
The ability of these agents to autonomously map environments, test for vulnerabilities, and adapt their strategies in real-time marks a significant evolution in threat actor capabilities. While the specific zero-day in MONAI Label has been mitigated, the underlying framework used to facilitate the attack remains a subject of concern for security professionals monitoring the intersection of artificial intelligence and cyber-offensive operations.






