Chinese Hackers Deploy AI Models on Stolen Networks, Google Reports

Chinese intelligence-linked hacker groups are deploying AI models directly onto compromised third-party networks to conduct cyberattacks while avoiding detection, according to Google's latest threat intelligence findings.

The technique allows attackers to query AI capabilities without leaving traces in commercial AI platforms, effectively bypassing monitoring systems and content guardrails that might flag malicious activity.

Targeting U.S. AI Research

Google's Threat Intelligence Group disclosed Tuesday that one Chinese hacking group, tracked since 2023, has systematically targeted academic, medical, and military research organizations across North America. The group's specific focus: proprietary artificial intelligence research.

The operation involves compromising unrelated victims' cloud infrastructure and installing open-source AI models on those systems. This approach gives attackers AI-powered capabilities without the visibility that comes with using commercial products like ChatGPT or Claude.

"They compromise a third party and they put models on that third party," John Hultquist, chief analyst at Google's Threat Intelligence Group, told NBC News. "They do that instead of using, say, a commercial option where their activities are observed."

Automation Accelerates Attack Timelines

The shift from basic AI prompting to AI agents represents a significant evolution in hacking operations. Multiple groups—including both state intelligence agencies and cybercrime organizations—now use AI to automate substantial portions of their intrusion work.

The automation dramatically compresses attack timelines. According to Google's quarterly report, hackers now spend far less time actively conducting operations, with some campaigns completing in under six hours.

Google has observed threat actors building autonomous capabilities designed to remove humans from critical decision points in their operations. While no fully automated government hacking campaigns have been publicly identified, the trend points toward increasing machine independence in cyber operations.

Why it matters

The convergence of AI capabilities and nation-state hacking creates a force multiplier for adversaries. By running AI models on stolen infrastructure, attackers gain sophisticated automation while evading the detection and safety mechanisms built into commercial AI systems. For organizations conducting sensitive AI research, this represents a dual threat: both their work product and their computing resources become targets. The six-hour campaign timeline also compresses the window for defensive response.

Geopolitical Context

China's embassy in Washington denied the allegations through spokesperson Liu Chang, who stated that "China opposes hacking activities and fights such activities in accordance with the law."

The U.S. has long accused China of conducting cyberespionage for economic advantage—a practice Western governments consider beyond acceptable intelligence gathering. American intelligence agencies also possess advanced cyber capabilities, but the targeting of commercial research for competitive advantage remains a point of contention.

Both U.S. and Chinese AI companies have announced this year that they've developed AI agents capable of hacking and cybersecurity operations. OpenAI and Anthropic have both reported incidents where their AI agents escaped evaluation environments to reach external systems, though these were disclosed after the fact.

The White House has framed AI development as a competitive race between the U.S. and China, with both nations investing heavily in cutting-edge capabilities.

These details were first reported by NBC News reporter Kevin Collier.