Anthropic Says It Blocked State-Linked Scientists From Using Claude For Bioweapons Research

Anthropic, the AI company behind the Claude family of models, has disclosed that it disrupted several attempts by scientists to use its systems for research that could have aided the development of biological weapons. The disclosure came in a new threat-intelligence report published titled 'Detecting and Countering Misuse of AI'.

What did Anthropic find?

According to the report, the company detected state-linked scientists attempting to use its AI models for research that could support the development of biological weapons, cyberattacks, influence operations, and conventional weapons development. The activity covered in the report spans December 2025 to August 2026, and includes five case studies of potentially dangerous biological research, among them gain-of-function work on the chikungunya virus, a mosquito-borne disease that can cause months of debilitating pain. The New York Time reported this on this threat-intelligence first.

The company said the researchers involved had circumvented controls meant to keep users from restricted regions off its models, and worked to obscure the purpose of their research to get past its safeguards. Anthropic responded by banning the accounts linked to the activity.

The Chikungunya Case

The most detailed example cited involves a scientist who, in May, approached Claude for help drafting a grant application to fund research on the chikungunya virus. While such research can be a legitimate step toward a vaccine, Anthropic said the scientist appeared to be seeking to engineer mutations that would make the virus more harmful through repeated infection of live animals. The company added that the institutional affiliation behind the request was one of the reasons it leaned toward viewing the research as less than innocent. Other reporting on the same case noted that the plan involved a military institute, even though civilian scientists had written the actual grant application.

A problem AI companies can't fully solve

Anthropic was candid about the limits of its own detection systems. The company said a classifier cannot simultaneously enable scientific benefit and prevent harm, since the same biological knowledge that produces a vaccine can just as easily be turned toward a weapon.

The report also acknowledged that Anthropic did not find any concrete instance of a scientist attempting to use its technology for outright nefarious purposes, and that the flagged use cases may well have overlapped with legitimate scientific inquiry into vaccines or other treatments. Even so, the company said it chose to act on caution rather than wait for certainty, banning the accounts regardless.

Withholding names, citing risk of exposure

Anthropic declined to name the researchers, their institutions, their countries of affiliation, or the specific pathogens under study in most cases. The company said in its report that it does not assert the researchers intended harm, and that identifying them or their labs could expose those individuals to harm. Separately, the company noted that it suspected the researchers in each of the five cases were connected in some way to a government or military in a country it has restricted from using its services, or had taken deliberate steps to conceal their location and identity. Anthropic's terms of service bar usage from a small set of countries, including China, Iran and Russia.

Beyond biology: Cyberattacks and weapons development

The report was not limited to biological research. Anthropic said it had also identified misuse of its models by suspected Russian, Chinese and Iranian government-aligned actors for cyberattacks, surveillance of diaspora communities and dissidents, and the development of software for conventional weapons such as missiles and armed drones. On the biology side specifically, the company noted a mixed record for its safeguards - in one case, a researcher studying bird flu was steered toward its weaker models, a restriction Anthropic characterised as providing only clerical-level assistance. According to one account, the flagged individual then simply moved to a rival AI model within days of being blocked.