Google DeepMind Expands AI Bioresilience Program to Prevent Biological Threats
Google DeepMind and Isomorphic Labs are expanding a major bioresilience initiative designed to prevent the misuse of artificial intelligence in biology while accelerating outbreak detection, scientific research, and the development of medical countermeasures.
The two organizations revealed that the initiative has established more than 15 partnerships with government agencies, biosecurity institutions, national laboratories, and research organizations over the past 12 months.
Among the named collaborators are Lawrence Livermore National Laboratory, the United Kingdom’s AI Security Institute, the Coalition for Epidemic Preparedness Innovations, or CEPI, and the Francis Crick Institute.
The program reflects a growing concern surrounding the dual-use nature of advanced AI. Frontier systems such as Google’s Gemini models can help researchers analyze diseases, identify therapeutic targets, design proteins, and accelerate vaccine development. However, the same capabilities could potentially assist individuals attempting to create or modify harmful biological agents.
DeepMind and Isomorphic Labs said their approach is built around a dual responsibility: enabling legitimate scientific breakthroughs while preventing advanced biological knowledge and AI tools from being exploited for malicious purposes.
Three Pillars of AI Bioresilience
The initiative is organized around three major areas:
- Preventing the misuse of AI in biology
- Detecting outbreaks and biological threats earlier
- Accelerating responses during outbreaks or biological attacks
The companies plan to expand their partnerships over the next six to 12 months, focusing on threat intelligence, AI agent evaluations, jailbreak prevention, and improved methods for handling high-risk biological training data.
DeepMind is also working with the Frontier Model Forum to explore how AI developers should manage sensitive datasets, including virology and pathogen-related information.
Preventing Harmful Use Without Blocking Scientific Research
DeepMind said its prevention strategy begins with threat modelling, which evaluates the types of individuals or organizations that may attempt to misuse AI and identifies the technical barriers currently limiting them.
The company uses expert-led red-teaming exercises and randomized controlled trials to assess whether models such as Gemini could help users overcome those barriers.
Post-training safety techniques are then applied to teach the models to reject harmful biological requests. At the same time, developers are attempting to avoid excessive refusal, where the system blocks legitimate questions from scientists, doctors, or researchers.
DeepMind also uses classifiers, internal probes, real-time monitoring, and targeted log analysis to identify suspicious activity that automated safety systems may fail to detect.
However, the company acknowledges that these protections remain a continuing area of research. Safety measures that work against known jailbreak techniques may not always stop new or previously unseen attack methods.
AI-Designed DNA Could Challenge Existing Screening Systems
One of the most serious concerns highlighted by the initiative involves synthetic DNA.
Members of the International Gene Synthesis Consortium currently screen customer orders by comparing requested DNA sequences against databases of known pathogens and toxins.
According to DeepMind, this approach may become less effective as AI systems gain the ability to design new DNA sequences that perform functions similar to dangerous biological agents without closely matching their known genetic sequences.
Such AI-designed sequences could potentially avoid existing screening mechanisms.
DeepMind is exploring whether SynthID, its watermarking technology for AI-generated images and text, could be adapted to identify or mark biological sequences created using AI.
The concept remains experimental and has not yet been deployed as a commercial or operational system.
A more advanced solution would involve developing screening technology capable of predicting whether an unfamiliar DNA sequence is toxic or pathogenic based on its likely biological function, even when the sequence does not resemble anything contained in existing databases.
DeepMind described this as a long-term technical challenge.
Lower-Cost Sequencing Could Improve Outbreak Detection
The detection component of the initiative focuses heavily on metagenomic sequencing.
Unlike conventional diagnostic tests that search for a limited set of known pathogens, metagenomic sequencing analyzes all microorganisms found in a biological or environmental sample.
This could allow authorities to identify emerging diseases before they spread widely. However, the technology remains expensive, particularly for countries and communities with limited healthcare and laboratory resources.
DeepMind pointed to a collaboration between Google and Pacific Biosciences that used the AlphaEvolve coding agent to improve sequencing accuracy.
The company is now exploring how AI could optimize data-processing algorithms, support sequencing hardware design, and help scientists understand new pathogens directly from their genetic sequences.
DeepMind is also investigating whether AlphaGenome, its genomic AI system, could assist researchers in characterizing unfamiliar pathogens.
These projects remain primarily research collaborations. Building a functioning early-warning network across airports, transit hubs, wastewater systems, hospitals, and densely populated areas would require significant infrastructure, funding, and international cooperation.
AlphaFold Supports Vaccine and Treatment Research
The response pillar focuses on closing the medical countermeasure gap.
Many known pathogens still lack approved diagnostics, vaccines, or treatments. DeepMind believes AI systems such as AlphaFold could help scientists understand disease-related proteins and accelerate the development of new medical interventions.
Over the past five years, more than 10,000 infectious-disease research publications have referenced AlphaFold, according to DeepMind.
The technology has contributed to studies involving tuberculosis, malaria, Mpox, Nipah virus, and other infectious diseases.
DeepMind has also partnered with Lawrence Livermore National Laboratory’s bioresilience program, which plans to use AlphaFold 3 to support the design of broad-spectrum antibodies.
This includes research into a possible pan-filovirus antibody that could target multiple viruses within the filovirus family.
The company said it will continue adding protein structures and molecular complexes to the AlphaFold Protein Structure Database, with priority given to targets relevant to infectious-disease treatments and emergency response.
Selected researchers are also being given access to newer AI research systems, including Google’s AI Co-Scientist.
Among the participating scientists are researchers from United States Department of Energy national laboratories working under the Genesis Mission.
Isomorphic Labs Forms Rapid Outbreak Response Unit
Isomorphic Labs, DeepMind’s sister company focused on AI-powered drug discovery, has established a dedicated unit that could rapidly deploy its drug-design platform during a new disease outbreak.
The unit is expected to work with governments, national laboratories, public health agencies, and scientific institutions.
Potential collaborators include Lawrence Livermore National Laboratory, CEPI, the UK AI Security Institute, and the Francis Crick Institute.
Isomorphic Labs has also pledged $7 million to Health for Human Potential, an initiative under the Philanthropy Asia Alliance, to support infectious-disease research across Asia.
The investment is particularly relevant to the region, where densely populated cities, frequent international travel, biodiversity, and unequal access to healthcare can increase the risk of emerging infectious diseases.
DeepMind Calls for Stronger US Biosecurity Policies
DeepMind also presented several recommendations to United States policymakers.
For prevention, the company supports the establishment of a federal safety framework for frontier AI systems. It also backed proposed legislation related to biological data standards, DNA synthesis screening, and the scaling of biotechnology infrastructure.
These proposals include:
- The AI-Ready Bio-Data Standards Act
- The Biosecurity Modernization and Innovation Act
- The SCALE Biology Act
For outbreak detection, DeepMind recommended expanding metagenomic sequencing across transportation hubs, urban centers, and other high-risk locations.
It also called for additional funding for early-warning research through agencies such as the Defense Advanced Research Projects Agency and the US Department of Health and Human Services.
For outbreak response, the company supports legislation designed to improve the sharing and integration of biological data.
DeepMind also recommended maintaining pharmaceutical manufacturing facilities in a ready state so they can be activated quickly during a health emergency. Other proposals include pre-established clinical trial networks and faster regulatory pathways for emergency diagnostics, vaccines, and treatments.
Many of the proposed measures still require legislative approval and government funding.
A Global Test for Responsible AI
DeepMind’s bioresilience initiative highlights the difficult balance facing the AI and biotechnology industries.
The same systems that could help scientists discover vaccines, understand pathogens, and respond to outbreaks could also lower the technical barriers for biological misuse.
The effectiveness of the program will depend not only on the performance of AI safety systems, but also on collaboration among governments, technology companies, public health institutions, researchers, DNA synthesis providers, and international organizations.
For the Philippines and other countries in Southeast Asia, the initiative also raises important questions about access.
Advanced AI-powered disease surveillance and drug-discovery systems may provide powerful protection against future outbreaks, but their benefits will remain limited unless developing countries receive access to the technology, infrastructure, training, and funding needed to use them.
AI may become one of the world’s most powerful tools for biological resilience. The challenge is ensuring that it protects humanity faster than it can be used to endanger it.
