Anthropic says Claude thwarted bioweapon research from state-sponsored actors — covert accounts used U.S. proxies to attempt to engineer deadlier viruses, tried to evade identification and regional blocks

In the span of five months from November 2025 to September 2026, Anthropic's AI Claude helped thwart potential bioweapon research attempts by state-sponsored actors. The company disclosed five instances where Claude was asked to assist in the creation of biological weapons, ranging from viruses to toxins. The primary challenge for Anthropic was determining whether the research was intended for beneficial purposes like vaccine development or nefarious goals such as enhancing virus virulence.

Three of the cases involved viruses, while the other two focused on toxins. In the first case, researchers sought assistance in creating a grant application to develop a chikungunya virus with increased mutation and virulence. The initial application appeared to be for civilian researchers, but it was later discovered that the actual research was intended for a military facility.

The request was routed through a third-party LLM platform associated with both military and civilian institutions, which used gray-market resellers and accessed Claude from the U.S. to bypass identification and regional blocks.

The second case involved a researcher who wanted to explore how avian flu adapts to mammals and can cause diseases beyond the respiratory tract. Although avian flu does not easily spread from person to person, understanding its potential to increase transmissibility could lead to dangerous outcomes. The researcher used a random username, a private email service, and accessed Claude through a Virtual Private Server (VPS), leading Anthropic to investigate and ultimately ban the account.

The third case resembled the first two, with another account attempting to prepare a grant application about orthopoxviruses, which include smallpox and Mpox. The application discussed containment facilities and live experimentation with the viruses, focusing on understanding their genetics to evade immunity. The research did not initially raise alarms, but Anthropic noticed it was generated via a reselling service, using a random email, tunneled through U.S. infrastructure, and traced back to a banned account farm.

Cases four and five dealt with toxins. In the fourth instance, a person mapped out venom toxin peptides from various animal families and created a program to optimize their toxic characteristics. The stated goal was to create painkillers, antidepressants, and other therapeutic molecules, but the data could equally be used to create potent harmful compounds.

The fifth case involved a theoretical scientist who tried to redesign a set of toxins, including a bacterial toxin subunit and a protein of a hemorrhagic-fever virus, for therapeutic purposes under a national public search program. However, the research touched upon a particularly nasty, pandemic-inducing disease as listed by the World Health Organization. Anthropic cut off access to Claude from a location that violated its terms of service.