Britain to launch world's first AI biosecurity lab

The United Kingdom has established the world's first AI biosecurity lab. According to a report by Hindustan Times, the lab has two primary objectives. First, to accelerate research using AI, improve the understanding of diseases, and assist in the discovery of new medicines and vaccines. Second, to ensure that AI is not used by any individual to create biological weapons.

The government will provide £100 million in funding for the lab, while companies OpenAI and Anthropic have made commitments of up to £200 million. Additionally, Google DeepMind will work with the lab for research and provide reliable access to its frontier Gemini models as well as specialised science models. This lab is part of the UK's broader £165 million National AI Biosecurity Initiative.

Key areas of bioscience AI research include molecular structures, involving tasks such as protein folding, genomic sequencing, and metabolic pathway mapping. Meanwhile, medicine AI research focuses on disease diagnosis, clinical trial matching, and predicting patient outcomes. Currently, collaboration between academia and AI companies is increasing, with two parallel streams of general-purpose frontier LLMs and specialised biological design tools (BDTs) being developed.

In August, Stanford University and the Arc Institute stated that generative AI has helped create complete and functional genomes for the first time. Using Evo 1 and Evo 2 genome language models, 16 designs were created that act as bacteriophages to infect and kill E. coli.

However, these advancements are accompanied by risks. In late February, Chen Bo and Kelvin Zhang, along with colleagues from Oxford University, UC Berkeley, and SecureBio, authored a research paper. They noted that Large Language Models (LLMs) are performing better on biology benchmarks, but it remains unclear whether they enable new users to outperform internet resources. Zhang, who was with Scale AI at the time, has been serving as a research program manager at OpenAI since June. OpenAI, Anthropic, and Google DeepMind are currently researching solutions to these problems.

The research organisation RAND warns that no single security measure can stop the development of weapons via AI. Sela Nevo, Director of the Center on AI, Security and Technology at RAND, stated, "We certainly do not know how quickly AI capabilities in biology will advance or which individual will attempt to misuse them, but we know that the window of opportunity to build preventive infrastructure is open now." RAND has suggested a multi-layered architecture, including access restrictions, precursor screening, misuse detection, and response, as an appropriate approach.

In general AI, capability increases with data and compute, but in biology, even a lightweight model fine-tuned on high-resolution genomic or structural data can design new toxins or new phages, as seen in the Stanford and Arc research.