Google DeepMind and Isomorphic Labs open their bioresilience playbook
The two AI labs say they will share models with vetted partners to speed vaccine design and spot outbreaks earlier, while adding safeguards to stop misuse.
Key points
- Google DeepMind and Isomorphic Labs published a joint bioresilience plan on July 16, 2026, covering prevention, detection and response to disease threats.
- The two labs say they have built more than 15 partnerships with governments, biosecurity groups and researchers over the past 12 months.
- DeepMind is adapting its SynthID watermarking system to biology, so DNA synthesis companies can flag AI-generated genetic sequences that look risky.
- Isomorphic Labs has set up a dedicated unit to rush its AI drug design engine to governments during novel outbreaks.
- The work sits inside DeepMind's Frontier Safety Framework, its internal rulebook for handling chemical, biological, radiological and nuclear risks.
Google DeepMind and its drug-discovery sister company Isomorphic Labs have laid out how they plan to stop their AI being used to build biological weapons, while offering the same technology to scientists trying to prevent the next pandemic.
The joint plan, published this week, calls the effort "bioresilience." In plain terms: keep bad actors away from the models, and put the good tools in the hands of people fighting outbreaks.
It matters because the same AI that can design a life-saving drug can, in theory, help design something dangerous. The labs are trying to widen the first door and lock the second.
What are they actually offering?
Access, in a controlled way, to some of the most powerful AI systems in biology today.
That includes AlphaFold, the DeepMind system that predicts the 3D shape of proteins (the molecular machines that run every living cell). It also includes AlphaGenome, which tries to explain what stretches of DNA actually do, and Isomorphic Labs' Drug Design Engine, an AI tool for designing new medicines.
Trusted researchers will be able to use these to speed up vaccine and treatment design for both known diseases and new ones nobody has seen yet.
For detection, DeepMind is pointing an AI agent called AlphaEvolve, software that can improve its own algorithms, at the maths behind metagenomic sequencing. That is the technique of scooping up all the genetic material in a sample (say, sewage or a nasal swab) and reading it at once to spot new pathogens. Cheaper analysis means more places can afford to watch for outbreaks.
How do they plan to keep the models from being misused?
DeepMind describes a four-step safety loop: threat modelling, evaluations, mitigations and monitoring. In practice that means in-house biologists and outside experts try to break the models, and the findings feed back into new guardrails on chatbots like Gemini.
The most interesting piece is a biological version of SynthID, the watermarking technology Google originally built to tag AI-generated images and text. Adapted to DNA, it would let the companies that physically print genetic sequences check whether an order came from an AI, and whether it looks dangerous, before they ship it.
That is a genuine chokepoint. Almost no serious biological work happens without ordering DNA from a synthesis provider first.
What does this mean for ordinary people?
Nothing changes on your phone tomorrow. But if the plan works, two things quietly improve.
First, health authorities may catch strange new pathogens weeks earlier than they do today, because sequencing gets cheaper and smarter. Second, when an outbreak does hit, the design of vaccines and antivirals could start days after the genome is read, not months.
The caveat is trust. DeepMind and Isomorphic Labs are choosing who gets access, what safeguards apply, and what "trusted partner" means. None of that is public in detail yet. Independent biosecurity researchers will want to see the evaluation results, not just the summary.
What to watch for as a reader: whether DNA synthesis companies actually adopt the SynthID biology checks, whether the partner list is published, and whether the Frontier Safety Framework's red-team results for biological risk are ever released in full.
Credit to the DeepMind and Isomorphic Labs safety teams for putting the shape of this program on the record. The proof will be in the audits.
