Anthropic Blocks Suspected Attempts To Use AI For Biological Weapons Research

The company disclosed the cases in its latest security and misuse report, saying it suspended the accounts involved after detecting behavior consistent with efforts to obtain sensitive biological research assistance. Anthropic said it could not determine with certainty whether the users were conducting legitimate scientific work or pursuing malicious objectives, but acted because of the potential risks.

The findings were published in Anthropic's latest transparency report on misuse of its Claude family of AI models and were first reported by The New York Times.

Jacob Klein, a senior security official at Anthropic, said distinguishing between legitimate biomedical research and potentially dangerous activity remains one of the company's biggest challenges.

“Research that can contribute to vaccines or medical breakthroughs can also be used to engineer more dangerous pathogens,” Klein said, adding that users with malicious intent rarely reveal their true objectives and often attempt to conceal them behind seemingly legitimate scientific queries.

One incident identified in May involved a researcher using Claude to draft a funding proposal for experiments involving the chikungunya virus. According to Anthropic, the proposal included plans for genetic modification of the virus and testing on animals, research commonly described as gain-of-function work, which has both legitimate scientific applications and significant biosafety risks.

Anthropic said the case became more concerning after investigators determined the proposed research was linked to a military research institution. The company did not identify the institution or the country involved, saying it could not verify whether the research was intended for defensive or offensive purposes.

The company also reported that several users attempted to bypass Claude's safety guardrails by disguising requests, using indirect prompts and concealing the intended purpose of their research. Anthropic said its monitoring systems detected these patterns and terminated the associated accounts.

The report warns that newer generations of AI models have become significantly more capable of assisting with complex biological analysis than earlier systems, increasing the importance of stronger safeguards around sensitive scientific information.

Susan Monarez, a microbiologist and public health expert who reviewed the report before publication, said the findings demonstrate that some individuals and organizations are actively exploring how AI could accelerate research into dangerous pathogens.

Andrew Weber, a former U.S. Defense Department official and senior fellow at the Council on Strategic Risks, described the report as a warning sign and argued that access to the most advanced AI tools for biological research should be limited to trusted researchers working under appropriate oversight.

Beyond biological research, Anthropic said it detected other forms of state-linked misuse of Claude. According to the report, actors associated with China and Iran used the chatbot to support surveillance of dissidents and expatriate communities, while Russian state media used it to generate online propaganda and fabricated election-related narratives about Moldova.

Anthropic also said it identified attempts to use Claude to assist with the design of conventional weapons, including firearms, rockets, drones and explosive devices. The company reported three such cases linked to users in China, two in Russia and one connected to Yemen, though it did not publicly identify the individuals involved.

The disclosure comes as AI companies face increasing scrutiny from governments and security experts over the potential misuse of increasingly powerful generative AI models. Anthropic says it continues to strengthen monitoring systems, restrict access to high-risk capabilities and cooperate with researchers and policymakers to reduce biological and national security risks associated with advanced AI.

MENAFN11092026000228011069ID1111651263