Scientists attempted to create biological weapons using Claude — Anthropic

2 hours ago

Anthropic announced that it has detected attempts to use Claude for cyberattacks, propaganda campaigns, and dangerous biological research.

The accounts of scientists who used artificial intelligence models "in ways that could contribute to the creation of biological weapons" have been blocked. Along with these cases, Anthropic also provided information about other threat actors who used AI for malicious purposes. Among them are criminals, likely state-sponsored groups, spyware vendors, and state propaganda entities.

Anthropic reported a total of five cases of attempts to create biological weapons. The company noted that these cases "indicate that the relevant capabilities exist" in AI models, but "cannot clearly demonstrate that they would be used to create biological weapons in the real world."

The individuals involved in these incidents were professional scientists. Anthropic decided not to disclose their names. They also stated there was no basis to say that these individuals aimed to harm humanity.

"When we identified and investigated these cases, we blocked the user accounts. We incorporated the findings of the investigation into our advanced protection systems, control mechanisms, and threat analysis processes to prevent, detect, and stop such activities in the future," the report says.

New and improved models like Claude Fable 5 have stronger protection mechanisms than previous models. It restricts access to "a wide range of queries related to dual-use biological research." At the same time, AI capabilities in the field of biology can be used for both beneficial and harmful purposes.

"The same information that can be used to develop biological weapons can also be applied, for example, to create a vaccine or a drug against a disease," Anthropic explained.

Propaganda and surveillance

Anthropic also announced that it had identified and deleted accounts that used Claude to shape public opinion. Among them were three accounts linked to Iranian state entities.

According to the report, in each of these cases, the operation was carried out by an individual working in or on behalf of an Iranian state propaganda institution. Claude was used to generate content, present posts as materials from independent news sources, and distribute them on social networks such as X, Instagram, and TikTok.

"One of the Iran-linked threat actors used Claude to generate recommendations on targeting US Navy assets in the region," Anthropic reported.

The same account also used AI to develop internal software for a mass surveillance platform intended for state systems in Iran.

Claude was also used by actors in China and West Africa to create and conduct surveillance operations, and in some cases, to identify targets. Anthropic also identified cases where AI practically replaced a team of engineers.

In China, three accounts linked to municipal security services used Claude for "surveillance and transnational repression." Among them was a municipal office that compiled profiles of activists and organizations abroad.

This is not the first case of an attempt to use artificial intelligence to create biological weapons. After being updated last summer, ChatGPT provided hundreds of users worldwide with instructions on creating and using biological weapons and toxins.