Anthropic halts AI help for risky virus grant

Anthropic, an artificial intelligence company, has halted AI assistance for a scientist's grant application concerning research that could potentially make the chikungunya virus more transmissible and better able to evade immunity, leading to a wide-ranging investigation into hazardous biological work tied to military and state-backed organizations.

The AI giant disclosed this incident in a September threat-intelligence report, which detailed five instances where researchers either utilized or attempted to utilize Claude, Anthropic's AI assistant, for activities that could support biological weapons development. Anthropic emphasized that it had not determined malicious intent and kept the scientist's identities, institutions, and countries anonymous due to the legitimate nature of biological research for medical advancements.

The chikungunya request, detected in May, involved identifying mutations, engineering them into infectious virus clones, and testing variants for properties such as transmissibility, immune evasion, and virulence. The proposed application suggested civilian researchers were involved, but the work was intended to be conducted at a military research institute, according to Anthropic.

Their biological safety classifier blocked exchanges related to the grant-writing request. Subsequent investigation revealed that the researcher accessed Claude via a platform serving life-sciences users in regions where Anthropic does not provide its services. The platform routed traffic through US infrastructure and employed methods to conceal users and circumvent regional restrictions.

Anthropic discovered that the platform's customers included virologists associated with both civilian and military institutions. They subsequently banned associated accounts, collaborated with partners to disrupt relay networks, and shared information with affected AI companies and government authorities. The disclosure is significant as it demonstrates that older Claude models were deemed insufficient for assisting sophisticated biological research, while newer systems possess sufficient capability that Anthropic no longer provides such assurances.

The episode highlights that active laboratory programs involving potentially enhanced pathogens are attempting to gain access to frontier AI systems while concealing their identities. Anthropic did not confirm whether the chikungunya project was a biological weapons program, acknowledging that similar experiments can contribute to vaccines, therapies, and understanding virus evolution.

Chikungunya is a mosquito-borne viral disease causing fever and severe joint pain, with most patients recovering although joint pain can persist for weeks, months, or longer in vulnerable individuals. Due to its natural circulation in many regions, a deliberate release could be difficult to distinguish from a natural outbreak. Anthropic noted that some of the activities proceeded because safety classifiers face challenges in advanced biology as the same technical knowledge can support both medicine and harmful applications.

They have progressively strengthened protections as their models have improved, implementing stronger deployment controls when systems may materially assist individuals seeking chemical or biological weapons, and employing stronger biological safeguards for newer, more capable models under certain conditions.