In September 2026, Anthropic released a threat intelligence report summarizing cases where its AI model "Claude" was misused. The report details multiple instances of exploitation, including cyberattacks, surveillance, biological risks, and model capability exploitation (distillation).

Regarding distillation, the report points to large-scale activities by Chinese AI development companies. Alibaba is reported to have conducted massive extractions exceeding 151 million times between May and July 2026. Additionally, Moonshot AI reportedly utilized proxy services to secretly forward customer requests to Claude to obtain responses. The report also cites cases where companies such as DeepSeek, Zhipu, and Xiaomi extracted inference processes (Chain-of-Thought).

In the field of cyberattacks, the report demonstrates how AI is being used to automate and accelerate attacks. It reports instances where state-sponsored or state-affiliated groups have leveraged AI for vulnerability research, malware development, phishing, and the construction of surveillance systems. These targets include government agencies, diplomatic missions, journalists, and specific ethnic groups.

Regarding involvement in biological risks, the report cites cases where AI was used for pathogen research, surveillance, and political influence operations. These include instances where AI assisted in studying pathogen mutations, enhancing infectivity, or supporting specific research. Anthropic stated that it has taken countermeasures, such as account suspensions, in response to these abuses and will continue to strengthen defensive measures to improve model safety.

Source: