Google DeepMind and Isomorphic Labs are expanding a bioresilience programme intended to capture AI’s scientific benefits without making biological misuse easier. Over the past year, they have formed more than 15 partnerships spanning government, biosecurity and research organisations.
The initiative is organised around three connected tasks: preventing harmful use, detecting outbreaks earlier and accelerating the response. Each exposes how far useful research has progressed—and how many safeguards remain experimental.
Prevention must distinguish dangerous intent from real science
Frontier models such as Gemini increasingly understand biology, while specialised models, agents and databases make that knowledge more actionable. The same capability that helps a researcher identify a vaccine target could help a malicious user fill gaps in a harmful plan.
DeepMind uses expert red-teaming and randomised controlled trials to test whether Gemini allows people to overcome known technical bottlenecks. Post-training aims to make models refuse dangerous requests without blocking legitimate scientific questions. Classifiers, internal probes and targeted log analysis look for risky behaviour during use.
None of these controls is presented as complete. A classifier that catches known jailbreaks under evaluation may miss a novel technique in production. Overly broad restrictions can also impede scientists working on pathogens, countermeasures or preparedness. The programme’s value depends on continuous testing with both security specialists and legitimate users.
DNA synthesis offers a concrete example. Current providers screen orders against known harmful pathogens and toxins. AI may enable the design of sequences that preserve dangerous function while differing enough from recorded sequences to escape a similarity-based check.
DeepMind is exploring whether an approach inspired by its SynthID watermarking system could mark AI-designed biological sequences. This is research, not a deployed safeguard. The more ambitious goal—predicting dangerous function from a novel sequence even when it resembles nothing in a database—remains an open technical challenge.
Detection depends on cheaper sequencing and field infrastructure
Metagenomic sequencing can identify all microorganisms in a sample instead of testing only for a shortlist of expected pathogens. That makes it attractive for early warning from wastewater, transit hubs and other population-level sources.
Cost is the principal obstacle, particularly in regions where outbreaks may begin but laboratory capacity is limited. DeepMind points to work between Google and Pacific Biosciences in which the AlphaEvolve coding agent improved sequencing accuracy. Further research may optimise data-processing algorithms, influence hardware design and use AlphaGenome to characterise pathogens.
These are meaningful building blocks, yet an improved laboratory pipeline is not an operating surveillance network. Real deployment needs sample collection, trained personnel, reliable supply chains, secure data exchange and protocols for deciding when an unusual signal warrants action.
Detection also raises governance questions. Metagenomic data can reveal information beyond the immediate public-health purpose. Systems need limits on retention and access, clear jurisdictional responsibility and safeguards against stigmatising communities when an alert appears.
DeepMind says it will widen partnerships over the next six to twelve months, focusing on threat intelligence, evaluation of AI agents and stronger jailbreak mitigation. Named collaborators include Lawrence Livermore National Laboratory, the UK AI Security Institute, CEPI and the Francis Crick Institute.
Faster response needs institutions as much as models
Thousands of infectious-disease papers have used AlphaFold, including work involving tuberculosis, malaria, Mpox and Nipah. A partnership with Lawrence Livermore will apply AlphaFold 3 to broad-spectrum antibody design, while selected public researchers are gaining access to systems such as Co-Scientist.
Isomorphic Labs has established a unit intended to deploy its drug-design platform rapidly during a novel outbreak. It is also supporting infectious-disease research in Asia with a $7 million commitment. These efforts target a genuine weakness: many known pathogens still lack licensed diagnostics, vaccines or treatments.
Models can narrow targets and propose candidates, but response speed also requires warm manufacturing capacity, pre-arranged clinical trial networks, regulatory pathways and purchasing commitments. DeepMind’s US policy recommendations call for those foundations alongside DNA screening, bio-data standards and wider metagenomic surveillance.
The cited legislation has not been enacted, and partnership announcements do not yet form a national biosecurity system. Success should be measured through adversarial evaluations, field detection times and countermeasures that reach trials—not merely the number of collaborators or model-generated hypotheses.
Bioresilience is therefore an institutional design problem powered by AI, not a model feature. DeepMind’s three pillars are sensible because no single safeguard can carry the burden. Prevention reduces opportunity, detection shortens uncertainty and response limits harm. The test is whether those layers work together under real pressure, when data is incomplete and decisions cannot wait for a perfect answer.
