Bioterrorism and AI: A Growing Threat to Global Security

AI and Bioterrorism: A Global Alert

Artificial intelligence is now capable of designing and conducting tens of thousands of biological experiments autonomously. This scientific advancement could, if misused, accelerate the development of deadly pathogens. Experts are sounding the alarm over this potential diversion.

Since the emergence of generative AIs, their ability to create various types of weapons has become an increasing concern. These technologies provide access to knowledge that was previously inaccessible to most individuals. OpenAI and Anthropic, for example, are seeking experts in chemical weapons to better control their models. Stephen D. Turner, an associate professor of data science at the University of Virginia, published an analysis in The Conversation, highlighting the growing danger of AI-facilitated bioterrorism.

Dual Use: A Double-Edged Sword

AI has made its way into biology laboratories, as it allows for rapid anticipation of molecular behaviors, where humans take years. The technology enables the design of new proteins, essential for the vital functions of our cells, in just a few hours. This helps researchers develop drugs more quickly and produce vaccines at a lower cost, thereby increasing the capacity to respond to epidemics. However, this revolution presents a risk of dual use.

The same tools can be repurposed for malicious ends. Specialists refer to this as the "dual use problem." Recent studies are concerning. Research conducted by Scale AI and the NGO SecureBio, which specializes in biosafety, showed that individuals without biology training, using a large language model, successfully completed complex virology tasks four times more often than a group without AI assistance. Even more worrying, nearly 90% of these users managed to bypass the security filters of the models to obtain sensitive information about dangerous pathogens.

Insufficient Safeguards

Safeguards are struggling to keep pace. The Biological Weapons Convention, signed in 1975, does not mention artificial intelligence. American companies that manufacture and sell synthetic DNA sequences must ensure that their products do not fall into the wrong hands, but this relies on their goodwill, without any legal obligation.

Moreover, safety assessments of AI models often lack transparency and are not designed to evaluate actual biological risks. Some companies, like Anthropic, are attempting to take action, but these initiatives remain voluntary and fragmented. Dario Amodei, CEO of Anthropic, acknowledged that the rapid development of AI could soon outpace companies' ability to assess its dangers, highlighting the magnitude of the challenge.

Understanding the Dual Use Problem

The dual use in AI-assisted biology describes a technology that can serve both beneficial purposes, such as medical research and vaccine development, and malicious uses, such as the design of pathogens. AI lowers the entry barrier by accelerating access to methods, hypotheses, and protocols, sometimes beyond the level of a non-specialist. Managing this dual use requires technical controls, such as filters and model assessments, as well as organizational measures, such as procedures and audits.

The Impact of AI on Experimental Biology

AI models can quickly propose candidate protein sequences and predict certain properties, such as structure, stability, and interactions. This significantly reduces the "idea → test" cycle. Coupled with automated platforms, they allow for thousands of trials to be launched in parallel. This gain is substantial in research and development, but the same capacity for rapid iteration can be diverted to explore risky avenues if controls are insufficient.

Security Filters of Language Models

The security filters of large language models combine rules, moderation patterns, and alignment techniques to refuse or frame certain sensitive requests. Their limitation is that they primarily evaluate the form and apparent intent of the request, while dangerous information can be obtained through rephrasing or fragmented requests. To mitigate this risk, it is necessary to test the models with realistic scenarios, measure their ability to produce actionable instructions, and enhance detection across query chains rather than on a single isolated question.