Saturday 21 November 2026
Security

Chinese AI agents found and exploited a MONAI Label zero-day

Amazon says the suspected Chinese operators' Cairn agents worked autonomously for nearly a day before using the flaw to gain root access to a university server.

Amazon Spheres from Sixth Avenue in Seattle
The Amazon Spheres seen from Sixth Avenue in Seattle, Washington. Photograph: Joe Mabel/Wikimedia Commons (CC BY-SA 4.0)

Amazon Threat Intelligence found a publicly hosted Cairn instance on September 10 that exposed a campaign against universities in East Asia outside mainland China, CISO CJ Moses said in a report shared with The Register. The suspected Chinese-linked attackers used the open-source pentesting framework to coordinate AI agents that were instructed to "obtain shell access" and left to operate on their own, BleepingComputer reported.

The campaign ran for 48 hours beginning on September 8, with prompts written in Simplified Chinese and script timestamps set to China Standard Time. Before the agents found the zero-day, they spent about 16 hours and 148 actions trying to compromise a Redis server by abusing its replication feature, escaping its scripting sandbox, writing to configuration files, and corrupting memory; none of those attempts succeeded. They also tried a Log4Shell injection, a server-side request forgery attack against a Qdrant instance, and credential sprays against PostgreSQL, SSH, and other services. A separate operation against a university GPU server ran for more than six hours and compiled 80 real usernames from public websites, academic papers, and other documents. BleepingComputer reported that the agents tried more than 30,000 credential combinations but failed.

"Neither operation succeeded, yet together they show that the agents kept adapting for nearly a full day without human direction," Moses wrote.

The agents then discovered a path traversal zero-day vulnerability in MONAI Label, an open-source application for medical imaging that uses AI. The agents determined the application's framework, enumerated its API endpoints, and located an upload endpoint with a filename parameter that permitted path traversal, Moses wrote. Within about 11 minutes, they went from mapping the API to writing arbitrary files on the server. Subsequently, the agents exploited the vulnerability to write a Python reverse shell into the sitecustomize.py file, which loads automatically when Python initializes, and then invoked an API endpoint to obtain root privileges. Later, a human operator gave the agents a proxy that connected to a second university network hosting the same vulnerable application.

Amazon informed the MONAI Label maintainers, who issued a patch and published a security advisory. BleepingComputer said the update was released on November 18. According to BleepingComputer, Amazon assessed with moderate confidence that the campaign was conducted by a Chinese-nexus threat actor, citing the language used in prompts, script timestamps, and target selection. The company shared two IP addresses observed between September 8 and 10: 142.4.98.145 and 142.4.107.34.

The Cairn abuse follows other autonomous AI agent incidents. Five Nines reported this month that an AI framework called exploit_agent, linked to a likely Russian-speaking operator, turned security patches into working exploits without human involvement, according to a Google Threat Intelligence Group report. Five Nines also reported that California sued OpenAI over experimental agents that attacked Hugging Face in July. Amazon's Moses wrote that the incidents indicate the autonomous AI agent framework is used in both espionage and financially motivated operations.

In separate Chinese-linked activity, a suspected Chinese APT exploited a VMware vCenter zero-day for more than two months before Broadcom patched it on Thursday, SecurityWeek reported. The flaw, tracked as CVE-2026-123890 with a CVSS score of 9.8, resides in the VMware Directory Service and allows remote, unauthenticated attackers to create an authenticated session with administrator rights. Mandiant linked the exploitation to Armory Castle and said the group used a tool called Bonesaw to steal the Lightning Memory-Mapped Database, extract SAML signing material, and create an SSO administrator account. Broadcom released fixes in vCenter versions 9.1.1.0200, 9.0.2.0200, and 8.0 U3m, but version 7.0 has no public patch because it reached end of general support in October 2025.

Sources: theregister, BleepingComputer, SecurityWeek, Five Nines (1, 2)

Lena Vasquez

Lena Vasquez covers vulnerabilities, breaches and identity for Five Nines. She writes from advisories, vendor bulletins and published reporting.

Corrections. Report an error to corrections@thefivenines.com. We list substantive corrections on the corrections page.