Google says China-linked hackers ran open-weight AI on compromised clouds

Google’s threat intelligence team says a China-linked group used victims’ cloud computing resources to run open-weight AI models while targeting North American research.

By Stat Crux Report1 min read

Published September 8, 2026 at 11:02 AM ET

The Googleplex headquarters in Mountain View, California. — Asoundd / Wikimedia Commons · Image source

Google’s Threat Intelligence Group said Tuesday that adversaries are moving beyond simple chatbot prompts toward agentic artificial-intelligence workflows that can automate portions of cyber operations. In one case described in its September 8 report, attackers compromised cloud resources and then planned, built and executed a mass credential-harvesting operation in less than six hours.

The report linked one campaign to UNC6508, a group Google assesses is connected to the People’s Republic of China. Google said the group has targeted North American academic, medical and military research organizations, including projects involving proprietary AI models, source code and API credentials.

According to Google, suspected UNC6508 operators also compromised cloud environments to deploy local, open-weight large-language-model infrastructure. Running the models on victims’ computing resources could help the operators avoid the monitoring and safety controls attached to commercial AI services. Google did not identify the affected organizations, and it said it has not yet observed threat actors deploying fully autonomous attack pipelines against targets in the wild.

A spokesperson for China’s embassy denied the accusations in comments reported by NBC News, saying China opposes hacking and rejects what the embassy described as smears. Google previously reported in June that UNC6508 had targeted research facilities in the United States and Canada between September 2023 and November 2025.