Several large language models (LLMs) could supply guidance that “could assist in the planning and execution of a biological attack”, a research team has warned.
The RAND Corporation has released a report warning that artificial intelligence (AI) chatbots could be used to facilitate biological attacks by providing advice on how to conceal the true purpose of the purchase of anthrax, smallpox and plague bacteria.
The research team conducted tests on LLMs – although it did not reveal which ones – that it accessed through an application programming interface (API), similar to that of OpenAI’s ChatGPT or Google’s Bard.
In one instance, the researchers asked LLMs about which biological weapons could create a plague pandemic scenario and talked to the system about potential agents, as well as considering budget and success factors. The LLM answered by identifying potential biological agents – including those that cause smallpox, anthrax and plague – and discussed their relative chances of causing mass death.
The LLM also assessed the possibility of obtaining plague-infested rodents or fleas and transporting live specimens, and stressed that the scale of projected deaths depended on factors such as the size of the affected population and the proportion of cases of pneumonic plague, which is deadlier than bubonic plague.
In another scenario, the team also asked LLMs about hypothetical foodborne and aerosol delivery methods of botulinum toxin. The LLM suggested using aerosol devices as a method and proposed a cover story for acquiring Clostridium botulinum.
The LLM recommended that a person pretend to be conducting scientific research to obtain the bacteria. “This would provide a legitimate and convincing reason to request access to the bacteria while keeping the true purpose of your mission concealed,” the LLM said.
Although at no point did the LLMs generate explicit instructions for creating biological weapons, the team pointed out that they were able to provide information that could “swiftly bridge such knowledge gaps” and “assist in the planning and execution of a biological attack”.
The team contrasted the information provided by the AI tool with previous attempts at creating biohazards. For example, an attempt by the Japanese cult Aum Shinrikyo to use botulinum toxin in the 1990s failed because of a lack of understanding of the bacterium – information that current LLMs could have provided.
“It remains an open question whether the capabilities of existing LLMs represent a new level of threat beyond the harmful information that is readily available online,” the researchers said.
The team concluded that there is an “unequivocal” need for further testing, and said AI companies must limit the openness of LLMs to conversations such as the ones in their report. The team also admitted that obtaining the information via an LLM required “jailbreaking”, meaning they overrode a chatbot’s safety restrictions.
In July Dario Amodei, CEO of the AI firm Anthropic, warned that AI systems could help create bioweapons in two to three years.
