Opinion: AI is eroding the barriers that kept biological weapons rare

Earlier this month, concerns arose that Anthropic's language models might be used to develop biological weapons. The company disclosed five instances where individuals leveraged their AI tools for potentially harmful purposes. One case involved assistance in preparing a proposal for gain-of-function research on a mosquito-borne virus at a military institution.

Another instance focused on planning experiments aimed at enhancing bird flu's ability to infect mammals. Although Anthropic's systems were unable to ascertain the true intentions behind these endeavors, they promptly intervened, blocking the activities in question. This ambiguity poses a significant challenge. The very tools that could aid scientists in crafting effective vaccines could also assist in creating more virulent pathogens.

Similarly, resources that enable the development of drugs targeting cancer cells while leaving healthy cells unharmed can be repurposed to construct medications that selectively target healthy cells. There is no requirement for someone to explicitly instruct Claude to "construct a bioweapon." It suffices for the individual to appear as a scientist engaged in crucial biological research.