Anthropic reported that this year it thwarted several potential attempts to use its AI models for research that could aid in the creation of biological weapons. The company published a report on 10 September 2026, in which it described cases where scientists tried to circumvent restrictions.
Researchers from prohibited regions bypassed access blocks and concealed the goals of their work. Anthropic could not determine precisely whether this research was legitimate or malicious — many biological experiments have dual use and can serve both for vaccines and for dangerous pathogens.
In one case, in May, a scientist sought help in writing a grant application for research on the chikungunya virus, which is transmitted by mosquitoes and causes severe pain. According to the company, the work was planned at a military research institute.
"You don't see someone in comic-book style saying, 'Hey, I want to build a biological weapon to kill everyone,'" noted Jacob Klein, head of Anthropic's threat division. The company chose to err on the side of caution and blocked the related accounts.
Expert Andrew Weber of the Council on Strategic Risks called these cases "alarming examples" of how state-linked biological weapons developers are using the capabilities of advanced AI models. The report covers five such incidents.
How far can such duality go in the real world, as AI becomes ever more powerful?
This is the first public report by a major AI company on real attempts at such misuse in the biological sphere, rather than only in laboratory tests. Anthropic emphasizes that the consequences of an error in assessment may be too serious to risk.
