Anthropic said it has neutralized several cases of alleged malicious use of its Claude models over the past eight months, including a cyberespionage operation linked to Russia and attempts by Chinese AI companies to extract and copy Claude's capabilities.
Anthropic said it has neutralized several cases of alleged malicious use of its Claude models over the past eight months, including a cyberespionage operation linked to Russia and attempts by Chinese AI companies to extract and copy Claude's capabilities.
In its latest Threat Intelligence report, Anthropic noted that cybercriminals and state-sponsored hackers are increasingly using AI not just for ancillary tasks, but to orchestrate and execute the bulk of cyberattacks. At the same time, humans are more likely to act as controllers than direct executors, writes Reuters.
“The use of AI has gone beyond simple chatbot questions and answers — it has involved engaging multi-agent systems to perform complex tasks,” Anthropic said.
Anthropic said it had stopped attacks from seven labs in China, including tech giant Alibaba, as well as developers Moonshot, DeepSeek, and Xiaomi.
According to the report, operators affiliated with Alibaba conducted the largest “illegal distillation” attack. They sought to extract the capabilities of Claude models to improve their own Qwen models. From May to July 2026, Anthropic recorded more than 151 million data exchanges from more than 3,500 fake accounts, with a peak load of almost 3 million requests per day.
Distillation is the process of training smaller, cheaper AI models using the responses of larger, more powerful models to reduce development costs.
At the same time, the developer of the Kimi chatbot, Moonshot and DeepSeek, instead of mass automated queries, redirected real dialogues of their users, which sometimes contained confidential information, through Claude, using his responses as training data.
Recall that the US authorities have accused Chinese AI companies of industrially copying American models . Among them are DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI.
The hacking group, whose style matches that of the Russian hacking center Midnight Blizzard (linked to the Russian Foreign Intelligence Service), conducted phishing operations, intercepted hotel Wi-Fi, and hijacked WhatsApp accounts against targets in the Ukrainian government, military, and diplomatic sectors.
The hackers used AI at almost every stage of the attack. In particular, they used AI to create a system that automatically detected when their malware was blocked by antivirus systems and rewrote the code until it was undetectable again.
Anthropic also discovered new categories of attackers who used Claude to develop software for conventional weapons, including:
firearms, missiles and armed drones;
ammunition and guidance and control systems;
intelligence gathering and procurement for military programs in China, Russia, and Yemen.
In addition, the company identified and terminated the activities of hackers associated with the ShinyHunters cybercrime group, which is behind a series of high-profile attacks on large corporations around the world.
Jacob Klein, head of threat analysis at Anthropic, emphasized that models have become significantly more powerful over the past year, which has created new risks: “A year ago, if you wanted to optimize a drone or software on a rocket, the models simply weren’t as adept at performing those tasks as they are now.”
