Anthropic Blocks Russian and Chinese AI Exploitation Campaigns

Anthropic has disrupted a series of state-backed cyber campaigns and intellectual property theft operations targeting its Claude AI models over an eight-month period. The operations involved Russian military espionage and Chinese AI labs attempting to weaponize the models for hacking, surveillance, and biological research.

According to reports from TechGolly and Nikkei Asia, the disruption uncovered a dual-threat landscape. Russian hacking group Midnight Blizzard utilized Claude to target more than 20 government agencies, embassies, and defense contractors across Ukraine and Europe. Simultaneously, Anthropic blocked five separate attempts to misuse the AI for biological research, including efforts to study the chikungunya virus.

The Industrial Scale of Model Theft

Beyond direct weaponization, Anthropic identified a systemic effort by Chinese firms to harvest the reasoning capabilities of Claude through "distillation attacks." This process involves using a frontier model to generate high-quality data that is then used to train smaller, domestic models.

Confirmed targets of these attacks included major Chinese firms such as Alibaba, DeepSeek, Moonshot, and Xiaomi. The scale of these operations was significant; Alibaba-affiliated networks generated over 151 million automated prompt exchanges between May and July alone to improve its Qwen model series, according to TechGolly.

A New Era of Automated Warfare

These events signal a critical shift in the nature of cyber warfare. Frontier AI models are now being used to create "automated evasion loops," which allow state actors to re-engineer malware in a matter of hours to bypass modern security software. This capability transforms the speed of offensive operations, moving from manual coding to AI-driven iteration.

Furthermore, the use of distillation demonstrates that the value of AI is not just in the service provided, but in the underlying "intelligence" that can be stolen. By harvesting proprietary reasoning, state-backed entities can accelerate the development of their own AI systems without the massive R&D costs typically required for frontier-level models.

Bypassing Global Restrictions

While Anthropic restricts access to its models in adversary nations including China, Russia, and Iran, the company noted that bad actors are increasingly sophisticated in their evasion tactics. Attackers frequently use fraudulent accounts and multi-agent frameworks to bypass geographic blocks and safety filters.

Industry analysts suggest that as AI capabilities grow, the battle between AI safety guardrails and state-sponsored exploitation will intensify. The focus now shifts to whether AI labs can develop detection mechanisms fast enough to counter automated, large-scale distillation and weaponization attempts before the proprietary intelligence is fully replicated.