Google DeepMind and Isomorphic Labs have announced the outline of a bioresilience program to curb the misuse of AI in biology while supporting outbreak response.
The organizations announced an update on a joint initiative that began in secret and has forged more than 15 partnerships with government agencies, biosecurity organizations and research groups over the past 12 months.
Certain framework issues are attached to this disclosure. DeepMind acknowledges that frontier models such as Gemini have an increasingly detailed understanding of biology, and that combining these systems with specialized biological models, agents like the Antigravity platform, and third-party databases will further enhance their capabilities.
However, the same knowledge that helps researchers map vaccine targets can, in principle, help threat actors close gaps in their own understanding. DeepMind and Isomorphic describe this as a double duty. This means enabling the cutting-edge scientific advances made possible by AI, while also ensuring that those same tools do not fall into the wrong hands.
The companies say the program is based on three pillars: preventing abuse, quickly detecting outbreaks, and responding if an outbreak or attack is in progress.
More than 15 partnerships built last year involve all three, but the latest information provides limited details about which organizations are involved beyond a few named collaborators, including Lawrence Livermore National Laboratory, the UK Institute for AI and Security, CEPI, and the Francis Crick Institute.
DeepMind says it intends to expand these relationships over the next six to 12 months, with a focus on threat intelligence, AI agent evaluation methods, and jailbreak mitigation. We are also working with the Frontier Model Forum on issues such as how to handle training data in high-risk categories, using virology datasets as an example.
Blocking Gemini without blocking legitimate science
Prevention efforts are based on threat modeling designed to identify which attackers are most likely to attempt exploits and what bottlenecks are currently stopping them. DeepMind says it is using a combination of expert red teams and randomized controlled trials to determine whether Gemini can help eliminate these bottlenecks.
The post-training method aims to teach the model to reject harmful queries while avoiding the company over-rejecting legitimate scientific questions. This balance has proven difficult, not just at DeepMind but across the industry. Classifiers and probes are being deployed to flag risky activity in real time, and the company says it is performing targeted log analysis to catch more subtle patterns of misuse that automated filters may miss.
None of these mitigations are listed as resolved. DeepMind frames them as continuous processes rather than finished systems. This is important for businesses and government agencies evaluating whether to rely on currently configured safety measures. Classifiers tuned against known jailbreak patterns in controlled evaluations do not guarantee equivalent performance against new attack techniques in the wild, and the company makes no claims to the contrary.
DNA synthesis screening problem
One of the more specific risks under investigation involves DNA synthesis. Companies within the International Gene Synthesis Consortium are currently screening orders against a list of known harmful pathogens and toxins in conjunction with screening algorithms. DeepMind has made it clear that this approach is starting to fray. Because AI can now help design DNA sequences that function similarly to dangerous pathogens without matching the sequences so closely as to trigger existing screens.
The proposed fix borrows from DeepMind’s existing watermarking system, SynthID, which the company says has become the industry standard for marking AI-generated images and text. Adapting it to biological sequences is presented as exploratory research rather than a shipped product.
Longer-term goals, described as unresolved technical challenges rather than close to a solution, include screening to predict whether novel DNA sequences have the potential to be toxic or pathogenic based on their function, regardless of whether they are similar to existing databases.
Cheaper sequences as detection layer
Detection relies on metagenomic sequencing, which characterizes all microorganisms in a sample, rather than checking a candidate list of known pathogens as in traditional diagnostics. The limiting factor is cost, which must be significantly lowered to expand the approach to areas where outbreaks are most likely to occur.
DeepMind points to a collaboration with Google and Pacific Biosciences to use the company’s AlphaEvolve coding agent to improve sequencing accuracy as one data point toward that goal. The company said it is currently considering further opportunities, from optimizing algorithms to process sequence data to informing hardware design, and is independently investigating whether AlphaGenome can help characterize pathogens directly from sequence data.
These remain research collaborations rather than field-deployed systems, and there is a long way to go to improve sequencing accuracy in controlled pipelines and operational early warning networks across wastewater and transportation hubs in low-resource settings.
AlphaFold’s publication record and countermeasure gap
Pillars of the response rely on gaps in medical preparedness, leaving many known pathogens without licensed diagnostics, vaccines, or treatments. DeepMind cites more than 10,000 publications on infectious diseases that referenced AlphaFold over five years, covering research on tuberculosis and malaria transmission and target mapping of threats including Mpox and Nipah.
The latest addition to this record is a partnership with Lawrence Livermore’s BioResilience Program, which plans to use AlphaFold 3 for broad-spectrum antibody design work, including work on panfilovirus antibodies. DeepMind says it will continue to add protein structures and complexes to the AlphaFold protein structure database this year, prioritizing relevant targets for countermeasure development.
Access to the new agent system, including Co-Scientist, is being expanded to select researchers, including scientists at U.S. Department of Energy National Laboratories working under the Genesis mission.
Isomorphic Labs has gone a step further, working with government and national research institutions such as Lawrence Livermore, the UK AI Security Institute, CEPI, and the Francis Crick Institute to establish a dedicated unit aimed at rapidly deploying drug design engines during new outbreaks. The company also pledged to donate $7 million to Health for Human Potential, an Asian philanthropy program, for infectious disease research across Asia.
DeepMind’s recommendations to U.S. policymakers directly address its three pillars and are based on specific pending legislation.
Regarding prevention, we support the federal Frontier AI Safety Framework, the AI-Enabled Biodata Standards Act (HR 7907), the Biosecurity Modernization and Innovation Act (S. 3741), and the SCALE Biology Act (HR 8981) to mandate DNA synthesis screening. Regarding detection, the American Living Libraries Act (S. 4770) accordingly calls for the enactment of the Biodata Act (HR 9307/S. 4770) and investment in “warm-based” and rapid activation-ready manufacturing capabilities, alongside pre-established clinical trial networks and faster regulatory pathways.
No such legislation has been enacted, and the gap between companies’ policy wish lists and a functioning federal biosecurity framework will provide the program’s real test over the next six to 12 months.
See also: Neko Health raises $700 million to expand AI body scanning in the US
Want to learn more about AI and big data from industry leaders? Check out the AI & Big Data Expos in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other major technology events such as Cyber Security & Cloud Expo. Click here for more information.
AI News is brought to you by TechForge Media. Learn about other upcoming enterprise technology events and webinars.
