Putting AI inside biology changes the safety question. The concern is not only whether a model can answer a dangerous question, but whether models, agents, databases and lab tools can be connected into a practical workflow.
On July 16, Google DeepMind and Isomorphic Labs published a joint bioresilience approach. The program has three tasks: prevent misuse of models, detect new outbreaks earlier and accelerate vaccines or other medical countermeasures. It sits between AI safety and AI drug discovery.
“Everyone agrees we can't get this wrong,” Google DeepMind vice president of responsibility Helen King told Axios.
The guardrail starts before release
DeepMind says it has advanced more than 15 partnerships over the past 12 months with government bodies, biosecurity organizations and research groups. The next 6 to 12 months will focus on threat intelligence, AI agent evaluations and jailbreak mitigations.
The safety process has four steps: threat modeling, evaluations, mitigations and monitoring. The aim is to judge whether systems such as Gemini could help a bad actor cross real-world bottlenecks, not just whether a single answer looks risky.
Biosecurity is moving from content moderation to task-chain control.
SynthID may move into DNA
The most concrete technical idea is adapting SynthID watermarking to biological data. DNA synthesis firms usually screen orders against lists of harmful pathogens and toxins. AI makes that harder because it can generate different sequences with similar function.
DeepMind wants partners to identify AI-generated sequences and, eventually, use AI to predict whether a sequence is likely to be toxic or pathogenic even when it does not resemble a known threat. That would move screening from name-list matching toward functional screening.
Detection depends on cheaper sequencing
The detection side points to metagenomic sequencing from wastewater, air or patient samples. Instead of checking only for known pathogens, it can scan all microorganisms in a sample and spot unusual outbreaks earlier.
Cost is the hard part. DeepMind says AlphaEvolve has already been used with Pacific Biosciences to improve sequencing accuracy, and it is exploring algorithm optimization, AlphaGenome and protein-function tools for pathogen characterization.
Google.org’s USD 7 million contribution to the Health for Human Potential coalition adds an Asia angle: the funding supports infectious-disease research and pandemic preparedness, with AlphaFold and Google Earth AI among the tools.
Isomorphic becomes the response arm
For response, DeepMind plans targeted access for trusted researchers to newer AI systems such as Co-Scientist. Isomorphic Labs has set up a focused unit that can deploy its Drug Design Engine for medical countermeasures during natural outbreaks or risks linked to advanced AI misuse.
The company also names Lawrence Livermore National Lab, the UK AI Security Institute, CEPI and the Francis Crick Institute as possible partners for real-world deployment.
The test is now operational. Who counts as a trusted partner, whether DNA watermarking can reach synthesis screening, and whether public-health agencies can afford near-real-time sequencing will decide whether this becomes infrastructure or remains a safety research program.
Sources: Google DeepMind / Isomorphic Labs bioresilience materials, Axios, Google DeepMind Singapore national AI partnership, Philanthropy Asia Alliance, CocoLoop; checked the 15-plus partnerships, three program pillars, SynthID biology watermarking, AlphaFold / AlphaEvolve / IsoDDE use cases and the USD 7 million Health for Human Potential funding figure.