On September 10 (local time), US-based Anthropic released its threat intelligence report, "Detecting and countering misuse of AI: September 2026," which summarizes instances of misuse of its AI, Claude. The report categorizes incidents detected and blocked from December 2025 to August 2026 into seven areas: cyberattacks, influence operations, surveillance, fraud, biological misuse, conventional weapons development, and distillation.
A particularly noteworthy point in the report is the "distillation" conducted by Chinese AI companies. distillation refers to the act of extracting a model's capabilities without authorization to replicate them in another model. Anthropic pointed out that companies such as Alibaba, DeepSeek, and Moonshot AI have been conducting industrial-scale campaigns using fake accounts or stolen credentials. Additionally, it revealed an instance where Xiaomi used Claude's responses to generate training data for the development of its own models.
Cases involving influence operations and surveillance were also reported. These include integration into production lines for Russian state media, efforts by state agencies to suppress anti-government activities, and surveillance activities targeting Uyghur support activists and human rights organizations. Regarding biological misuse, five cases were reported involving the design of pathogens or research into toxins.
Anthropic stated that as model capabilities improve, risks are also increasing, and that they are strengthening restrictions for next-generation models.
Source: