Anthropic blocks attempts to access AI Claude for possible bioweapons research

Anthropic provides insight into alleged attempted abuse of its AI Claude. Scientists from blocked regions tried to secretly access AI models - for research projects that are at least questionable.

In recent months, the AI ​​company Anthropic says it has prevented several attempts by researchers to use various models of its Claude software for the possible development of bioweapons. At the same time, the competitor of the ChatGPT developer OpenAI said that the targets of such requests were difficult to determine. Corresponding research could serve both medical and military purposes.

In any case, the experiments showed that scientists from blocked regions and with connections to local government agencies were trying to secretly access artificial intelligence from American developers, it was said. Neither the research institutions nor the countries from which they come were named.

Anthropic detailed five cases related to possible bioweapons research in a detailed report on attempted misuse of its AI models. The researchers were not looking for specific scientific help, but rather organizational support, for example in writing papers. AI models are fundamentally designed to refuse assistance in the development of biological weapons. However, in research that can serve both medical and military purposes, it is sometimes difficult to draw a clear line between harmless and malicious intentions, Anthropic explained.

Risky research on bird flu

In one of the access attempts blocked by the AI company, an attempt was made to write an application for funding for research on a virus with the help of Claude. This was about so-called gain-of-function research, in which the transmissibility of a pathogen is increased. A common method in the search for antidotes in laboratories - but also a way to develop biological weapons.

In another case, an attempt was made, with the help of Claude, to design a plan for research into a bird flu variant that could be transmitted to mammals. Here too, Anthropic emphasizes that such research can serve as prevention, but at the same time creates dangerous knowledge that can be used in weapons development. Another scientist wanted to use Claude to prepare quarterly progress reports when researching a toxin.