Anthropic Says It Stopped 5 Cases of Claude Being Used in Bioweapons-Adjacent Research

Anthropic has published its most detailed misuse report yet, and the bio section is the part that should make people squint. Over roughly eight months, the company says it found five cases where scientists used Claude for biological research that could have supported work on biological weapons. Anthropic says it could not tell whether any of those projects were legitimate or malicious, so it chose caution over certainty. That’s the uncomfortable bit: the same methods used for vaccine research can also be used to design dangerous pathogens.

The examples are specific and unsettling. One involved a grant application on chikungunya virus that included changes meant to increase transmissibility and evade immune responses, and it was linked to a military research institute. Other cases covered weeks of planning and analysis around highly pathogenic avian influenza, plus work on orthopoxviruses, toxins and venom-related compounds. Claude’s protections blocked the sensitive requests and cut off access to more capable model versions. In at least one case, the people involved simply moved over to a competitor’s model with weaker guardrails.

Anthropic banned the accounts and shared the findings with authorities and other AI companies, but it did not name the institutions, countries or individuals. That omission is deliberate. The company says it does not know the intent behind the cases, and naming names could put people at risk. Jacob Klein, who leads threat intelligence at Anthropic, says these incidents are subtler than the cartoon version of bad behavior. Nobody just types in that they want to build a biological weapon.

The broader point is bigger than the bio scare. Over a 30-day review period, Anthropic says it found 35 research efforts from state-linked institutions, most of them legitimate civilian work, which is exactly why simple filters get messy fast. A system built to enable useful science also has to stop harmful science, and those goals collide.

The report also widens the frame to conventional weapons for the first time. Anthropic describes cases in China, Russia and Yemen, along with surveillance, propaganda and cyber operations from state-linked and criminal actors. The blunt takeaway is that AI has pushed down the skill threshold. The old rule was that serious operations needed serious teams. That rule looks much shakier now.