Artificial intelligence company Anthropic says it has blocked several attempts in recent months by researchers seeking to use its Claude software for the development of biological weapons.
The company, a competitor to ChatGPT developer OpenAI, said the purpose of such requests was often difficult to determine because the underlying research could serve both medical and military purposes. Anthropic said, however, that the attempts demonstrated that scientists from blocked regions and researchers with links to state institutions there had sought to gain covert access to artificial intelligence developed by US companies.
The company did not identify the countries or research institutions from which the requests originated.
In a detailed report on attempted misuse of its AI models, Anthropic identified five cases involving possible biological weapons research. According to the company, the researchers were not primarily seeking specific scientific assistance from Claude. Instead, their requests involved organisational tasks, such as assistance with writing papers.
AI models are generally designed to refuse assistance with the development of biological weapons. Anthropic said, however, that drawing a clear line between legitimate and malicious intentions can sometimes be difficult when research has potential applications in both medicine and military activities.
One of the blocked attempts involved using Claude to prepare a funding application for research into a virus. The proposed work involved so-called gain-of-function research, in which the transmissibility of a pathogen is increased.
Such research is commonly used in laboratories in the search for ways to counter pathogens, according to the supplied account. At the same time, Anthropic noted that the same type of research can also provide a route towards the development of biological weapons.
Another case involved an attempt to use Claude to develop a research plan for a variant of avian influenza that could be transmitted to mammals. Anthropic said such research could contribute to prevention efforts, while also creating dangerous knowledge that could potentially be used in weapons development.
The company also identified an attempt by another scientist to use Claude in preparing quarterly progress reports relating to research into a toxin.
The cases illustrate a particular challenge for AI companies seeking to prevent their systems from being used for harmful purposes. Research into pathogens and toxins can have legitimate medical objectives while simultaneously generating knowledge that could potentially be applied to biological weapons.
Anthropic’s report therefore highlights not only efforts to prevent direct assistance with weapons development, but also the difficulty of assessing seemingly administrative or organisational requests connected to sensitive scientific research.
The company said its findings showed attempts by researchers in blocked regions, including individuals with connections to state institutions in those areas, to access American-developed AI systems covertly. However, without identifying the countries or institutions involved, Anthropic provided no further details about the origins of the researchers.
The cases have added to concerns over how advanced AI systems might be used in highly sensitive areas of biological research, particularly where legitimate scientific objectives and potential military applications overlap. Anthropic’s response has been to block requests it considers connected to biological weapons development, while acknowledging that determining the intentions behind some forms of research remains difficult.
