Anthropic blocks possible attempt to use AI to make biological weapons

Anthropic has reported blocking attempts by scientists to use its AI models for biological weapons development over an eight-month period ending in August 2026. The company's threat intelligence report detailed five case studies where actors used Anthropic's Claude models in ways that could support such development, highlighting biological misuse as a serious risk of frontier AI. Anthropic also detected and disrupted other malicious uses of its technology, including cyber and influence operations, surveillance, scams, fraud, and weapon development by various groups, including suspected state-sponsored actors. The company accused Chinese AI firms of attempting to replicate Claude's capabilities. These disruptions occurred across Anthropic's Claude Haiku, Sonnet, and Opus models, with one instance of model distillation also noted.

American AI company Anthropic has accused Chinese AI rivals DeepSeek and Moonshot of secretly routing user requests to Anthropic's Claude models and presenting the responses as their own. This accusation escalates US criticism of Chinese AI industry practices. The allegations follow claims by US government agencies, including the NSA and FBI, that six Chinese companies are involved in stealing US model outputs, potentially with Chinese government knowledge. Beijing has denied these allegations, calling them an attempt by Washington to suppress China's AI industry and threatening retaliation.

Anthropic, a US technology company, has reported disrupting several attempts to use its AI models, including the chatbot Claude, for research that could potentially aid in the development of biological weapons. The company's report, released on Thursday, details misuse cases over the past eight months, highlighting concerns about the potential dangers of AI development. Anthropic stated it erred on the side of caution when it could not definitively determine if research inquiries were legitimate or malicious, citing that vaccine research could also inform the engineering of dangerous pathogens. One specific case involved a scientist seeking assistance with a grant application to engineer mutations to the chikungunya virus to make it more harmful in live animals, with the work potentially intended for a military research institute. Experts are urging for stricter access to AI models, suggesting they should be limited to trusted researchers, as newer AI models are capable of complex scientific work that older versions could not perform.