AI vs. Biology: A Dangerous Game ⚠️🧬

BIO-RESILIENCE PROGRAM: MITIGATING AI MISUSE IN BIOLOGY
DeepMind and Isomorphic Labs have launched a bioresilience program designed to curb the misuse of artificial intelligence in biological research while simultaneously bolstering outbreak response capabilities. This initiative, initiated quietly over the past 12 months, has already fostered partnerships with more than 15 government bodies, biosecurity organizations, and research groups, reflecting a proactive approach to managing the potential risks associated with increasingly sophisticated AI models.

A DUAL MANDATE: SCIENTIFIC ADVANCE AND SAFETY
Frontier AI models, such as Google’s Gemini, demonstrate an increasingly detailed understanding of biological systems. DeepMind acknowledges that combining these models with specialized biology tools, like its Antigravity platform, and third-party databases will amplify their capabilities. However, the same knowledge that facilitates vaccine target mapping could also be leveraged by malicious actors seeking to accelerate their own understanding of pathogens. DeepMind and Isomorphic Labs operate under a dual mandate: to enable the scientific advancements driven by frontier AI while simultaneously preventing its misuse.

THREE PILLARS OF THE PROGRAM
The bioresilience program is structured around three key pillars: preventing misuse of AI tools, accelerating outbreak detection, and facilitating effective response measures once an outbreak or attack occurs. The 15+ partnerships established over the past year align with these three pillars, though specific details regarding the involved organizations remain somewhat limited, with confirmed collaborations including Lawrence Livermore National Laboratory, the UK AI Security Institute, CEPI, and the Francis Crick Institute. DeepMind intends to expand these relationships over the next six to twelve months, focusing on threat intelligence, AI agent evaluation methodologies, and jailbreak mitigation strategies. The company is also coordinating with the Frontier Model Forum to address concerns surrounding the handling of riskier training data, specifically virology datasets.

ACTIVE MITIGATION AND CONTINUOUS MONITORING
DeepMind employs a multi-faceted approach to mitigate potential risks, utilizing expert red-teaming and randomized controlled trials to assess Gemini’s capabilities. Post-training methods are implemented to instruct the model to refuse harmful queries while avoiding over-refusal of legitimate scientific questions – a challenge prevalent across the AI industry. Real-time classifiers and probes flag risky activity, and targeted log analysis identifies subtle misuse patterns that automated filters might miss. DeepMind emphasizes that these mitigations are an ongoing process rather than a fully realized system, acknowledging that performance may vary depending on the context and evolving attack methods.

DNA SYNTHESIS SCREENING: A FRAGILE DEFENSE
One significant risk involves DNA synthesis. The International Gene Synthesis Consortium currently screens orders against lists of known harmful pathogens and toxins, supplemented by screening algorithms. However, DeepMind recognizes that AI can now assist in designing DNA sequences with similar function to dangerous pathogens without closely matching their sequences, rendering existing screens ineffective. The proposed solution involves adapting DeepMind’s SynthID watermarking system, a standard for marking AI-generated images and text, to biological sequences – a project currently in exploratory stages. A longer-term goal involves screening based on DNA sequence function, regardless of similarity to known pathogens, utilizing metagenomic sequencing to characterize microbial communities. Scaling this approach to regions with high outbreak risk requires substantial cost reductions.

ALPHAEVOLVE AND METAGENOMIC SEQUENCING
DeepMind’s collaboration with Google on its AlphaEvolve coding agent to improve sequencing accuracy, combined with the exploration of metagenomic sequencing, represents a potential pathway toward more effective outbreak detection. The company is also investigating further opportunities to optimize sequencing data processing and inform hardware design. A separate effort, exploring whether AlphaGenome could characterize pathogens directly from sequence data, remains a research collaboration. The distance between sequencing accuracy gains in controlled environments and functioning early-warning networks in resource-constrained settings is a considerable challenge.

ALPHAFOLD’S CONTRIBUTIONS TO COUNTERMEASURES
DeepMind’s AlphaFold publication record, encompassing more than 10,000 infectious disease references over five years, has contributed to research on threats such as tuberculosis and malaria. The latest addition involves a partnership with Lawrence Livermore’s bioresilience program, utilizing AlphaFold 3 for broad-spectrum antibody design, including a pan-filovirus antibody effort. DeepMind intends to continue expanding the AlphaFold Protein Structure Database, prioritizing targets relevant to countermeasure development, and extending access to agent systems like Co-Scientist to selected researchers, including those at the US Department of Energy’s National Laboratories.

INVESTMENTS IN RESEARCH AND COLLABORATION
Isomorphic Labs has taken a further step by establishing a dedicated unit to deploy its drug design engine rapidly during outbreaks, working alongside government and national research bodies. The company has committed $7 million to Health for Human Potential, a Philanthropy Asia Alliance program, supporting infectious disease research across Asia. DeepMind’s recommendations to US policymakers align directly with its three pillars, reflecting a proactive engagement with regulatory considerations.