· ai-devtools
AI agents autonomously find and exploit 0-day in MONAI Label to take over university server
The AI agents demonstrated remarkable self-adaptation, operating for almost a full day without any input from a human operator, according to Amazon CISO CJ Moses.
These autonomous AI agents, suspected to be employed by Chinese spies, successfully located and exploited a zero-day vulnerability within an open-source application, ultimately allowing them to breach university networks and seize complete control of a server. The malicious actor utilized Cairn, a penetration-testing framework available as open source, which has grown increasingly popular among cybercriminals for executing complex, multi-stage intrusions without direct human intervention when paired with AI tools. Specifically, these agents discovered and exploited a zero-day in MONAI Label, an open-source application designed for medical imaging AI. Following Amazon's notification to the project's maintainers regarding this weakness, a patch was issued, accompanied by a security advisory.
Moses detailed in a report shared with The Register on Friday that Amazon's Threat Intelligence Team identified a publicly accessible instance of Cairn on September 10, which provided insight into the ongoing campaign targeting universities across East Asia, excluding Mainland China. This campaign spanned a 48-hour duration, commencing on September 8, and the prompts used were written in Simplified Chinese, with script execution timestamps set to China Standard Time.
Prior to successfully finding and exploiting the zero-day, the AI agents dedicated approximately 16 hours to attempting infiltration of a different university's Redis server, executing 148 sequential actions. These attempts included leveraging Redis's replication capability to load unauthorized code, escaping its scripting sandbox, modifying its configuration files, and causing memory corruption. None of these earlier efforts succeeded, leading the agents to subsequently attempt a Log4Shell injection against a separate service, launch a server-side request forgery (SSRF) attack against a flaw in a nearby Qdrant instance, and perform several credential sprays against PostgreSQL, SSH, and other utilities. Concurrently, a separate operation ran for over six hours, targeting a university GPU server with another credential-spray assault while compiling a list of 80 genuine user names sourced from the school's public websites, academic papers, and other documentation. Moses observed, "Neither operation succeeded, yet together they show that the agents kept adapting for nearly a full day without human direction."
It was only later that the AI agents located the path traversal vulnerability within MONAI Label, which granted them server access on this occasion. Moses noted, "The agents identified the application's framework, mapped its API endpoints, and found an upload endpoint with a filename parameter." After testing this parameter for path traversal, the vulnerability was confirmed. Within roughly 11 minutes, the agents progressed from mapping the application's API to achieving the ability to write files arbitrarily onto the server. Subsequently, the agents exploited this same flaw to execute a Python reverse shell operating with root privileges. In a subsequent session, after the zero-day was uncovered, a human operator furnished the agents with a proxy that tunneled into a second university network, connecting to a server running the identical vulnerable code.
This discovery builds upon a previous finding by Gambit Security's threat intelligence team, which had documented a Chinese operator using Cairn to conduct at least 105 attacks, resulting in compromises, "to varying degrees," across a minimum of 27 corporations. This campaign also took place during September. Furthermore, beyond Cairn, the attacker employed two distinct open-source AI harnesses to pilfer credit card records and other sensitive information from entities such as a major US airline, a large private US industrial supplies distributor, a US online fashion retailer, and a Fortune 500 hospitality company. Moses concluded, "Together, these cases show that this autonomous AI agent framework is being adopted in both espionage and criminal operations."
Why it matters
This demonstrates a significant evolution in threat actor capabilities, moving toward fully autonomous exploitation chains. Developers must be aware that popular open-source tools are being leveraged by sophisticated actors for zero-day discovery and exploitation.





