DeepMind and Isomorphic Launch AI Bioresilience Program

Google DeepMind and Isomorphic Labs announced a joint bioresilience program aimed at preventing misuse of AI in biology, detecting outbreaks faster and accelerating countermeasure development. The companies say the initiative expanded quietly over the past 12 months and now includes more than 15 partnerships, among them Lawrence Livermore National Laboratory, the UK AI Security Institute, CEPI and the Francis Crick Institute.

The program rests on three pillars. The prevention pillar focuses on stopping misuse of biological knowledge generated by AI. The detection pillar aims to scale sequencing that can spot novel threats sooner. The response pillar targets faster design of diagnostics, vaccines and therapies.

For prevention, the partners use threat modelling to identify likely misusers and the obstacles they face. They apply expert red-team exercises and randomized controlled evaluations to test whether AI models remove those obstacles. Post-training techniques are being developed to make models refuse harmful queries without blocking legitimate scientific requests. The companies deploy real-time classifiers and probes to flag risky activity and run targeted log analysis to find subtler misuse patterns. The update notes these safeguards are a work in progress and not final solutions. The companies are coordinating with the Frontier Model Forum on how to treat higher-risk categories of training data, using virology datasets as an example.

A specific technical risk under review concerns DNA synthesis screening. Current industry practice screens orders against lists of known harmful sequences. The companies warn that AI can design sequences with similar harmful functions that do not closely match known entries and could evade those filters. DeepMind and Isomorphic are exploring adaptations of SynthID, a watermarking tool used for AI-generated content, to mark AI-designed biological sequences. They are also pursuing longer-term research into screening methods that predict whether a novel sequence is likely toxic or pathogenic based on function rather than sequence similarity.

Detection work focuses on metagenomic sequencing, which profiles all organisms in a sample rather than testing for a short list of pathogens. The companies identify cost as the main barrier to deploying metagenomic surveillance where outbreaks are most likely to start. DeepMind cites a Google–Pacific Biosciences collaboration that used an AlphaEvolve coding agent to improve sequencing accuracy and says it is exploring algorithm and hardware optimizations and testing whether AlphaGenome can characterise pathogens directly from sequence data. The partners describe these efforts as research collaborations that are not yet field-deployed.

On response, the update highlights AlphaFold’s use in infectious disease research, noting more than 10,000 publications have referenced AlphaFold across topics including tuberculosis and malaria and work on threats such as Mpox and Nipah. A Lawrence Livermore partnership plans to use AlphaFold 3 for broad-spectrum antibody design, including pan-filovirus efforts. Isomorphic Labs established a unit to deploy its drug-design platform during novel outbreaks and has extended access to select agent systems to researchers, including scientists at U.S. Department of Energy national laboratories under the Genesis Mission. Isomorphic also pledged $7 million to Health for Human Potential to support infectious disease research in Asia.

The companies provided policy recommendations to U.S. lawmakers that align with the three pillars. Those proposals include support for a federal frontier AI safety framework and backing for pending bills cited by the companies: the AI-Ready Bio-Data Standards Act (H.R. 7907), mandatory DNA synthesis screening through the Biosecurity Modernization and Innovation Act (S. 3741), the SCALE Biology Act (H.R. 8981), expanded metagenomic sequencing under the America’s Living Library Act (S. 4023), and the Web of Biological Data Act (H.R. 9307 / S. 4770). None of the cited legislation has been enacted.

DeepMind and Isomorphic plan to widen partnerships over the next six to twelve months with emphasis on threat intelligence, evaluation methods for AI agents and jailbreaking mitigations. The companies describe the program as ongoing research and list technical and policy gaps that they intend to address through further collaboration.