Anthropic Blocked 35 Attempts to Use Claude for High-Risk Biological Research

Anthropic Blocked 35 Attempts to Use Claude for High-Risk Biological Research

The company could not determine whether the research was legitimate or malicious — yet it cut off access anyway, including after a request for a gain-of-function study of chikungunya tied to a military research institute.

Anthropic identified 35 attempts to use its AI model Claude to accelerate high-risk biological research during a 30-day evaluation period, according to the company's threat report as covered by Mindstream. The report was announced Thursday, according to The New York Times, and lands amid a wave of AI risk warnings — including a weekend call from Anthropic's own founder Dario Amodei to slow down development.

The chikungunya case

One of the cases involved a researcher who asked Claude for help writing an application for gain-of-function research on the chikungunya virus, a mosquito-borne disease that can cause months of severe pain, as reported by The New York Times. The research was said to be connected to a military research institute.

Jacob Klein, Anthropic's head of threat intelligence, said according to the NYT that the company did not know whether the research was intended for weaponization, but found the combination of a military institute and gain-of-function research concerning.

What else the report shows

The report's five case studies concerned inquiries about bird flu, poison, toxins, transmissibility, and immune escape, according to Mindstream. Users circumvented location-based restrictions and gave Claude false information about their research goals to avoid the platform's restrictions.

The report also documented attempts to use the system to develop software for conventional weapons systems — firearms, missiles, armed drones, and bombs — with three cases in China, two in Russia, and one in Yemen, according to the NYT. The country figures apply to the weapons-system cases, not the bio cases, for which no identities, institutions, or countries were disclosed.

Uncertain intent, blocked anyway

Anthropic acknowledges that the company could not determine whether any of the research was legitimate or malicious, but shut it down because of the potentially severe consequences, according to the NYT. Connected accounts were banned. The figure of 35 refers to identified attempts over a 30-day period, not confirmed bioweapon actors, and the number of blocked accounts has not been reconciled against the number of attempts in the coverage.

Why now

Anthropic itself writes that older models, when evaluated last year, were not capable of providing meaningful assistance with dangerous biological research — but that today's models are significantly more capable of complex scientific work. That is the background for the tightened security measures, according to the company.

Former CDC director Susan Monarez, after reviewing the findings, warned that advanced AI increases the opportunities to "let bad actors hide in plain sight, using seemingly legitimate research to create pathogens we may not see coming, and that we may not be able to stop once they're released," as quoted via Mindstream.

Amodei followed up the following weekend with a post writing that it would be wise to ban "using AI for the production of biological weapons or allowing users to do so," as quoted by The New York Times.

Experts urge caution

According to the NYT, today's chatbots are, by expert assessment, unlikely to be able to help individuals construct a deadly pathogen. But newer models trained on biology may require stronger security measures, according to the same analysis.

Many questions remain open: who the flagged users were, which institutions and countries they belonged to, and how many accounts were actually banned have not been disclosed. Anthropic's primary report is not available in the accessible source material — all of the findings above are based on secondary coverage in The New York Times and Mindstream.