AI company Anthropic has revealed five instances of the use of its Claude models in biological research that could pose risks of misuse, including research related to modifying viruses and toxins. The company said it had blocked accounts associated with these activities and enhanced its protection systems after investigating them. According to a threat report from Anthropic.

Anthropic stated that the misuse of AI in the biological domain represents one of the most dangerous risks associated with advanced models, especially as their ability to assist in complex scientific tasks develops. It explained that the report covers activities it monitored and disrupted between December 2025 and August 2026, across seven different areas of misuse.

In one case, the company discovered an attempt to use Claude to prepare a research grant application involving studies on the Chikungunya virus with the aim of studying properties related to its transmission and ability to evade the immune response.

Anthropic said that some of the research materials used formulations that downplayed the nature of the modifications, leading the company to conclude that the activity was not limited to preparing a funding application.

In another case, the company monitored a researcher outside of the United States using “Cloud” in research related to highly pathogenic avian influenza and its adaptation to mammals. Anthropic indicated that the use was carried out through an intermediary serving more than 12 clients, and witnessed the exchange of tens of thousands of messages with the company’s models over a short period. According to an Anthropic report.

Other cases included research related to toxins and toxic substances; the company said that one researcher used “Cloud” to build a database for toxic peptides and develop a computer system to improve molecules targeting functions associated with paralysis and pain relief, while another researcher used the model in a computer redesign of a set of toxins within a national research program.

Anthropic did not disclose the names of the institutions or countries where these activities took place, nor did it specify the biological factors and technologies used in detail, explaining that the individuals mentioned in the cases are working in the field of scientific research, and that the company does not claim that they intended to cause harm.

The company said that some users attempted to bypass geographical restrictions and conceal the nature of their research to circumvent protection systems, and that it closed the accounts linked to these activities. It also used the results of its investigations to develop new systems for detecting and preventing attempts to misuse its models.

These findings come at a time when concerns are growing that artificial intelligence could be used not only to speed up scientific research, but also to facilitate dual-use activities that can serve legitimate civilian purposes or be exploited for harmful purposes. Anthropic emphasized that the cases it uncovered do not mean that Claude has already led to the development of a biological weapon, but rather they indicate the need for more sophisticated controls as model capabilities increase.

In a separate report, Reuters said that the cases uncovered by Anthropic included attempts to use Claude in research on viruses and toxins, along with other instances of using the model in cyber operations, surveillance, espionage, and fraud. The agency noted that Anthropic said it had stopped these activities and blocked the accounts associated with them.

• Anthropic
• Artificial Intelligence
• Biological Weapons