Google DeepMind and Isomorphic Labs:
Our approach to bioresilience
Google DeepMind has pioneered the application of AI to grand challenges in biology, from protein
structure prediction to gene expression. Isomorphic Labs is using AI to accelerate drug discovery. Our
joint AI bioresilience program brings these capabilities together. Today, we are sharing an update on the
program, which seeks to prevent threat actors from misusing our AI models, while ensuring that experts
can use the models to prevent, detect, and respond to future outbreaks.

Biosecurity risks are growing. High rates of deforestation, urbanization, and travel increase the risk of a
destructive natural outbreak. Fast-improving AI models and agents may erode some of the knowledge
barriers that help prevent threat actors from developing and deploying bioweapons.

We need the scientific advances that frontier AI models will enable to make society more resilient to these
biosecurity risks. But we need to achieve this while also putting in place stringent measures to prevent
threat actors from misusing AI. This is the dual mandate of Google DeepMind and Isomorphic Labs’ AI
bioresilience program.

The program has three pillars:
   ●​ Prevent threat actors from misusing our models.
   ●​ Detect new outbreaks quickly.
   ●​ Respond decisively to these outbreaks.

Over the past 12 months, we have advanced more than 15 partnerships with government bodies,
biosecurity organizations, and research groups across these three areas. We will advance these
partnerships further in the next 6-12 months. If you are interested in partnering, please reach out here.

1. Prevent threat actors from misusing our models
Frontier AI models like Gemini have an increasingly sophisticated knowledge of biology. As they are
integrated into agents like Antigravity, and paired with specialized AI biology models and third-party tools
and databases, their capabilities will become stronger.

To ensure that Gemini is safe, but still useful to scientists and experts, we use threat modeling to
understand which actors are most likely to be willing and able to carry out an attack. We draw on various
evaluation methods—from expert red-teaming to randomized controlled trials—to help judge whether
Gemini could help threat actors overcome the primary bottlenecks they face.

We deploy various mitigations to counter these risks. Our post-training methods teach the model to
identify and refuse to assist queries with harmful intent, while aiming to avoid the over-refusal of beneficial
science queries. Our classifiers and probes detect and block risky user activity, while our targeted
analysis of user logs helps detect more subtle patterns of misuse and ensures that our mitigations are
working effectively. These mitigation efforts are complemented by robust security infrastructure and
privacy controls.

We partner closely with government bodies, external experts, and peers across this risk reduction
process. In the next 6-12 months, our priorities include threat intelligence, evaluation methods for AI
agents, and jailbreak mitigations. We are also working with our peers at the Frontier Model Forum to align
on best practices for difficult questions, such as how to treat riskier datasets—for example in virology.
To prevent attacks, there are also specific risks that we are exploring further. For example, we want to
help ensure that AI models do not make it easier for threat actors to access DNA. Leading DNA synthesis
companies, like those in the International Gene Synthesis Consortium, use lists of harmful pathogens and
toxins, alongside screening algorithms, to help detect potentially risky orders. However, this approach is
starting to fray, as AI makes it possible to design different DNA sequences with similar functionality.

To help DNA companies identify AI-generated sequences, we are exploring how to adapt our
watermarking technology, SynthID, which has become an industry standard, to biological data. In the
future, we also hope to help partners use AI to develop a new kind of DNA synthesis screening that could
help predict the function of a DNA sequence and whether it is likely to be toxic or pathogenic, irrespective
of whether it resembles a known pathogen or toxin—a major, open technical challenge.

2. Detect new outbreaks quickly
To detect new outbreaks, the world needs ubiquitous biosurveillance that can sequence genomic
data—from wastewater, the air or patients—and characterize the pathogens present in it, in close to
real-time. Traditional approaches detect a small number of known pathogens. Metagenomic sequencing
can detect novel or less common outbreaks by sequencing all microorganisms in a sample.

To scale metagenomic sequencing to the parts of the world where it is most needed, it needs to become
much more cost-effective. A recent collaboration between Google and Pacific Biosciences used our
AlphaEvolve coding agent to improve the accuracy of sequencing and we are now exploring other
opportunities, from optimizing the algorithms used to analyze sequencing data to informing
hardware design. We are also exploring how technologies like AlphaGenome could be used to help
detect and characterize pathogens from sequence data.

3. Respond decisively to these outbreaks
A large number of known pathogens have no licensed diagnostic, vaccine, or treatment available, leaving
the world highly vulnerable to new outbreaks. Researchers are using our models to help close this
medical countermeasure gap.

Over the past five years, AlphaFold has been cited in more than 10,000 publications on infectious
diseases. Researchers have used it to help better understand tuberculosis and malaria transmission, and
to map vaccine and drug targets for threats like Mpox and Nipah. We also recently partnered with the
bioresilience program at Lawrence Livermore National Lab, which will use AlphaFold 3 to accelerate
broad-spectrum antibody design, such as the pan-filovirus antibody. To enable this kind of research, we
have worked with our partners to steadily add relevant protein structures and complexes to the AlphaFold
Protein Structure Database. We plan to add more this year, with a focus on targets for medical
countermeasure development.

We are also granting leading scientists targeted access to our most recent AI agent systems, such
as Co-Scientist, which helps researchers develop compelling hypotheses. This includes researchers in
the US Department of Energy’s National Laboratories, who we are collaborating with as part of the US
government’s Genesis Mission. These scientific partners are pursuing a range of goals, such as
identifying novel drug combinations for pathogens that are resistant to current treatments. To scale these
impacts, we recently pledged US$7 million in support for Health for Human Potential, led by the
Philanthropy Asia Alliance, for infectious disease research in Asia.

The complexity of biology means that it is impossible to predict, with perfect precision, the future
outbreaks or attacks that may occur. To support government bodies and non-profit organizations during
novel outbreaks, Isomorphic Labs has established a focused unit to rapidly deploy its drug design
engine to help develop medical countermeasures that could address both naturally occurring
pandemics and potential risks arising from the misuse of advanced AI. To ensure real-world impact, we
would deploy such capabilities in collaboration with leading governments and national research centers,
such as Lawrence Livermore National Lab, the UK AI Security Institute, CEPI, and the Francis Crick
Institute.

Our recommendations for policymakers
We are partnering with government bodies around the world across these three pillars. We also make the
following recommendations to US policymakers:

1. Prevent threats by passing a national frontier AI safety framework and modernizing DNA
screening
Establish a federal framework to address biosecurity and other national security risks from frontier AI
models, including guidance on how to judge risk-benefit trade-offs. Pass the AI-Ready Bio-Data
Standards Act (H.R. 7907) to ensure that biology training data is safe, standardized, and suitable for
training AI models. We also recommend making DNA synthesis screening mandatory, via the US
Biosecurity Modernization and Innovation Act (S. 3741), and directing NIH and NSF to carry out research
to evolve screening towards more robust structure- and function-based methods. We also recommend
passing the SCALE Biology Act (H.R. 8981) to help develop the measurement standards, guidelines, and
best practices needed to support safe biotechnology.

2. Detect biosecurity incidents quickly by expanding early-warning infrastructure
Identify and characterize new outbreaks before they spread more widely, by aggressively implementing
metagenomic sequencing across key locations like transit hubs and high-density areas. Support this by
passing the America’s Living Library Act (S. 4023) to sequence and catalog genomic data across public
lands, creating a database for biosurveillance. Agencies like DARPA and HHS should also scale up
funding for cutting-edge research into new kinds of AI-enabled early warning and attribution systems.

3. Respond decisively by accelerating medical countermeasure pipelines
Invest in adaptable diagnostic and therapeutic platform technologies that can be quickly customized to
counter new biological threats. Enable this by supporting the creation of secure, centralized, high-quality
biology datasets, via the proposed Web of Biological Data Act (H.R. 9307 / S.4770). Complement this
with investment in the infrastructure needed to quickly test, scale up, and deploy novel countermeasures,
for example by “warm-basing” and evolving manufacturing facilities so that they are operationally ready,
pre-establishing clinical trial networks and creating accelerated pathways for regulatory approval.

____