Google DeepMind、AI の生物濫用防止プログラムを発表

Google DeepMind and Isomorphic Labs outlined a bioresilience program to curb AI misuse in biology while aiding outbreak response.

The two organisations published an update on a joint initiative that began quietly and has now built out more than 15 partnerships with government bodies, biosecurity organisations, and research groups over the past 12 months.

The disclosure arrives with a specific framing problem attached. Frontier models such as Gemini carry an increasingly detailed grasp of biology, and DeepMind acknowledges that pairing these systems with specialised biology models, agents like its Antigravity platform, and third-party databases will only sharpen that capability further.

However, the same knowledge that helps a researcher map a vaccine target could, in principle, help a threat actor close gaps in their own understanding. DeepMind and Isomorphic describe this as a dual mandate: enable the scientific advances frontier AI makes possible, while keeping those same tools out of the hands of people who’d misuse them.

The program sits on three pillars, according to the companies: preventing misuse, detecting outbreaks faster, and responding once an outbreak or attack is underway.

The 15-plus partnerships built over the last year touch all three, though the update gives limited detail on which organisations are involved beyond a handful of named collaborators, including Lawrence Livermore National Laboratory, the UK AI Security Institute, CEPI, and the Francis Crick Institute.

DeepMind says it intends to widen these relationships over the next six to twelve months, with attention turning to threat intelligence, evaluation methods for AI agents, and jailbreak mitigations. It’s also coordinating with the Frontier Model Forum on questions such as how to handle riskier categories of training data, virology datasets being the example given.

Locking down Gemini without blocking legitimate science

The prevention work rests on threat modelling designed to identify which actors are most likely to attempt misuse and what bottlenecks currently stop them. DeepMind says it uses a mix of expert red-teaming and randomised controlled trials to judge whether Gemini could help someone clear those bottlenecks.

Post-training methods are meant to teach the model to refuse harmful queries while avoiding what the company calls over-refusal of legitimate science questions, a balance that’s proven difficult across the industry generally, not just for DeepMind. Classifiers and probes are deployed to flag risky activity in real time, and the company says it runs targeted log analysis to catch more subtle misuse patterns that automated filters might miss.

None of these mitigations is described as solved. DeepMind frames them as an ongoing process rather than a finished system, which matters for any enterprise or government body evaluating whether to rely on the safeguards as currently configured. A classifier tuned against known jailbreak patterns in a controlled evaluation doesn’t guarantee equivalent performance against novel attack methods surfacing in live use, and the company doesn’t claim otherwise.

The DNA synthesis screening problem

One of the more concrete risks under exploration involves DNA synthesis. Companies within the International Gene Synthesis Consortium currently screen orders against lists of known harmful pathogens and toxins, paired with screening algorithms. DeepMind states plainly that this approach is starting to fray, because AI can now help design DNA sequences with similar function to a dangerous pathogen without matching its sequence closely enough to trigger existing screens.

The proposed fix borrows from DeepMind’s existing watermarking system, SynthID, which the company says has become an industry standard for marking AI-generated images and text. Adapting it to biological sequences is presented as exploratory work, not a shipped product.

A longer-term goal, described as an open technical challenge rather than something close to resolved, involves screening that predicts whether a novel DNA sequence is likely toxic or pathogenic based on its function, regardless of whether it resembles anything in existing databases.

Cheaper sequencing as the detection layer

Detection depends on metagenomic sequencing, which characterises every microorganism in a sample rather than checking for a shortlist of known pathogens the way traditional diagnostics do. The limiting factor is cost, and scaling the approach to the regions where outbreaks are most likely to originate requires that cost to fall considerably.

DeepMind points to a collaboration between Google and Pacific Biosciences that used its AlphaEvolve coding agent to improve sequencing accuracy as one data point toward that goal. The company says it’s now looking at further opportunities – from optimising the algorithms that process sequencing data, through to informing hardware design – and separately exploring whether AlphaGenome could help characterise pathogens directly from sequence data.

These remain research collaborations rather than field-deployed systems, and the distance between a sequencing accuracy gain in a controlled pipeline and a functioning early-warning network across wastewater and transit hubs in low-resource settings is not small.

AlphaFold’s publication record and the countermeasure gap

The response pillar leans on the medical countermeasure gap that leaves many known pathogens without a licensed diagnostic, vaccine, or treatment. DeepMind cites more than 10,000 publications on infectious disease that have referenced AlphaFold over five years, covering work on tuberculosis and malaria transmission and target mapping for threats including Mpox and Nipah.

The newest addition to that record is a partnership with Lawrence Livermore’s bioresilience program, which plans to use AlphaFold 3 for broad-spectrum antibody design work, including a pan-filovirus antibody effort. DeepMind says it will keep adding protein structures and complexes to the AlphaFold Protein Structure Database this year, prioritising targets relevant to countermeasure development.

Access to newer agent systems, including Co-Scientist, is being extended to selected researchers, among them scientists in the US Department of Energy’s National Laboratories working under the Genesis Mission.

Isomorphic Labs has gone a step further, setting up a dedicated unit intended to deploy its drug design engine quickly during a novel outbreak, working alongside government and national research bodies such as Lawrence Livermore, the UK AI Security Institute, CEPI, and the Francis Crick Institute. The company also pledged $7 million to Health for Human Potential, a Philanthropy Asia Alliance programme, for infectious disease research across Asia.

DeepMind’s recommendations to US policymakers map directly onto its three pillars and lean on specific pending legislation:

On prevention, it backs a federal frontier AI safety framework, the AI-Ready Bio-Data Standards Act (H.R. 7907), mandatory DNA synthesis screening through the Biosecurity Modernization and Innovation Act (S. 3741), and the SCALE Biology Act (H.R. 8981).

On detection, it wants metagenomic sequencing expanded across transit hubs and dense population centres, supported by the America’s Living Library Act (S. 4023) and additional DARPA and HHS funding for early-warning research.

On response, it calls for the Web of Biological Data Act (H.R. 9307 / S. 4770) and investment in manufacturing capacity kept “warm-based” and ready for rapid activation, alongside pre-established clinical trial networks and faster regulatory pathways.

None of that legislation is enacted, and the gap between a company’s policy wishlist and a functioning federal biosecurity framework is where the real test of this program will play out over the next 6-12 months.

See also: Neko Health raises $700 million to expand AI body scans in the US

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.

AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.

The post Examining Google DeepMind’s AI bioresilience push appeared first on AI News.