Euroviews. Brussels should know AI-assisted bioterrorism is a risk worth considering

AI regulation has become an area of inter-state competition. While the EU just reached a deal on the AI Act, the US has previously released a far-reaching executive order on AI, and the UK convened political and industry leaders at the AI Safety Summit. 

In many of these discussions, one risk is getting more attention: AI-assisted bioterrorism, or the ability of individuals to cause catastrophe by using AI tools to get access to a pandemic virus.

We recently showed that this is a risk worth considering. In an informal experiment, we tasked individuals with using an open-source large-language model that had its safeguards removed to help them obtain a pathogen capable of causing a pandemic. 

Within three hours, participants identified many of the steps required to start a potentially catastrophic outbreak. 

Particularly concerning was that the model advised participants on how to access the viral DNA — the blueprint for creating the pathogen — while evading existing screening methods. The extent to which current models aid bioterrorism by summarising information that already exists online remains unclear. 

However, current capabilities aside, the findings suggest that in the absence of robust safeguards, more advanced future models might provide malicious individuals with streamlined and accessible information on how to access, construct, and release a pandemic virus.

The DNA constructs required to build a virus from scratch can be ordered online: many gene synthesis providers will manufacture a thousand base pair pieces of DNA for under €200 — something that only a few decades ago took researchers thousands of hours and warranted a Nobel Prize. 

The advent of custom gene synthesis is now a pillar of the biological sciences,