This lab will serve two key objectives. First, to use AI to accelerate research, improve disease understanding, and support the discovery of new medicines and vaccines; and second, to ensure AI doesn’t help anyone build a biological weapon.
A week earlier, the United Kingdom became the first country to set up an AI Biosecurity Lab, the first of its kind anywhere in the world.
Earlier, Google DeepMind announced it is working with Isomorphic Labs on a bio-resilience programme built around prevention, detection and response. Google DeepMind will also work closely with the Lab for research, and look to provide trusted access to frontier Gemini models as well as their specialised science models. The lab sits inside the UK’s broader £165m National AI Biosecurity Initiative and sees Silicon Valley’s fiercest rivals come together as the boundary between AI models, coding and physical biology disappears fast. Looking back to late February, when Chen Bo Calvin Zhang wrote in a research paper titled ‘LLM Novice Uplift on Dual-Use, In Silico Biology Tasks’ alongside colleagues from the University of Oxford, UC Berkeley and SecureBio, a single uncertainty sat at the heart of scientific acceleration and dual-use biosecurity risk. They are collaborating with governments, biosecurity groups and researchers. The pitch is clear—frontier AI is not just a biosecurity problem, it is a biosecurity tool. Anthropic had taken the most visibly restrictive line initially, and now seems most visibly uncomfortable with that. That means that it can provide genuine assistance to a researcher developing a new medical treatment (which is the reason we’re so keen to widen access to the model via both classifier improvements and trusted access programs). But in the wrong hands, those same capabilities could be used by a malicious actor, for example in developing a biological weapon,” the company explained, in the biology safeguards disclosure in August. In September, the company began to unlock more access for vetted life-science professionals. At that time. The company clarified that they had repeatedly shut down misuse attempts, including research that could have supported bioweapons work. The risk landscape In the current scenario, government regulations globally are meant to provide guardrails for governing general purpose AI and frontier models, not specifically biology usage scenarios. AI companies realise this. AI companies including OpenAI, Google DeepMind, Microsoft AI and Anthropic, as well as scientists and researchers, have urged the US government to pass legislation mandating universal screening and record-keeping for synthetic DNA sequences. legislative efforts are ongoing While this hasn’t been passed yet.
The AI Biosecurity Lab will have £100 million in government funding, and up to £200 million in company commitments from OpenAI and Anthropic. Its powerful Claude Fable 5 model launched with safety classifiers that were so aggressive they blocked plenty of ordinary biology questions, but the change of approach became clear with an update in August that cut those false blocks by about 85%. “One such risk is in the field of biology: Fable 5 can now outperform experts on some highly complex biological tasks and provide operational support on others. OpenAI has a $50,000 bounty programme that calls on vetted applicants to try and break its biosafety guardrails, which it would then be able to fix.
In August, Stanford University and the Arc Institute said that generative AI had helped compose a complete, functional genome for the first time, with 16 designs becoming bacteriophages that infected and killed E. coli., and was done using genome language models Evo 1 and Evo 2. On October 21, they confirmed that a collaboration with PerturbAI, and Scripps Research that used about 2,000 AI agents powered by GPT‑Rosalind to analyse and interpret existing experimental data from roughly 8 million mouse brain cells in just 32 minutes, in an attempt to find new clues about Alzheimer’s and Parkinson’s. Researchers noted that OpenAI’s GPT-5.5 made significant generational strides on complex genetics benchmarks like GeneBench. Meanwhile, Anthropic’s Mythos frontier model managed to solve 30% of complex bioinformatics problems on the BioMysteryBench evaluation that a panel of human domain experts could not.
The GPT-Rosalind life-sciences reasoning model was released in April, followed in late May with the Rosalind Biodefense programme that sponsors vetted developers building pandemic-preparedness tools and extends access to government partners. In October, Google DeepMind revealed that their Google Antigravity agentic science workbench can simplify the process of researching new therapeutics in the search for new antibiotics to counter rising levels of antimicrobial resistance. “We build cell atlases so that other scientists can make discoveries we haven’t imagined yet.
There is collaboration between academia, which is increasingly using specialised AI models, and AI companies. The key to this is two parallel streams of development—general-purpose frontier LLMs, and specialised Biological Design Tools, or BDTs. Research organisation RAND warns that no single safeguard can prevent an actor from using AI to try and develop weapons. They point to multi-layered architecture as a viable counter-approach: access restriction, precursor screening, misuse detection, and deterrence and response. In general AI, capability scales predictably with massive compute and data. In biology, a relatively lightweight model fine-tuned on high-resolution genomic or structural data can design novel phages as the Stanford and Arc research did, or toxins. “Developing a drug involves many varied tasks: doing early research to identify a target for a drug and the right type of drug to engage it, designing a molecule with the right properties, optimising it to be manufactured reliably at scale, drafting the regulatory filings needed to run clinical trials, designing and administering those trials, and much more. Last week, they detailed early stage research progression, including increasing the usefulness of a cancer treatment known as CAR T-cell therapy, developing drugs that would detach the IgE molecule that sets off a chain reaction leading to food allergies from human immune systems, and using AI to flag issues in clinical trials by scanning thousands of pages of documentation. Different paths, similar missions Each of the big three AI labs has moved forward significantly this year, yet each has picked a slightly different posture. OpenAI is focusing more on what it terms defensive acceleration. In June it published an action plan, ‘Biodefense in the Intelligence Age,’ arguing that societies must build resilience faster than AI raises risk. The approach is delivering results. Researchers at DeepMind point out that there has been exponential growth in genomic and metagenomic sequencing, which has populated global databases with billions of natural amino acid sequences. This pool of natural data has helped produce bacteriocins, alongside recent genome-mined breakthroughs such as Darobactin, Dynobactin, and Lariocidin. Global security organisation NTI, or Nuclear Threat Initiative, in their latest AIxBio Horizon Scan points out that protein design tools have continued to improve, agentic coding tools have matured in ways that lower barriers to computational biology, and commercial AI companies are making significant investments in the life sciences.
“We don’t know exactly how fast AI capabilities in biology will advance, or which actor will try to exploit them first, but we do know that the window for building preventive infrastructure is open now,” says Sella Nevo, director of RAND’s Center on AI, Security, and Technology. Our strategy to realise AI’s medical benefits is based on having AI increasingly involved in all of these tasks,” Anthropic says in a statement. AI could help more researchers explore these resources and turn unexpected observations into new questions to test,” says Sarah Teichmann, Ph.D., Chair in Stem Cell Medicine at the University of Cambridge, and founding co-chair of the Human Cell Atlas. “Many, however, remain locked within the unannotated “dark matter” of protein databases, with their function uncharacterised because they aren’t similar to known proteins. To bridge this gap, Google developed ProtNLM (Protein Natural Language Model), a deep learning model trained to directly read the underlying evolutionary “grammar” of amino acid sequences,” they say.
A few weeks ago, DeepMind researchers led by Pushmeet Kohli introduced SynthID Bio, a proof-of-concept watermarking system embedded directly into amino acid sequences and physically synthesised proteins. “Screening is also one of the best understood and least disruptive biosecurity measures available. It asks providers of synthesised DNA and manufacturers of synthesis machines to check synthesis requests for sequences of concern and to verify customer legitimacy before shipping orders,” the open letter states. Important signatories from the AI industry include Sam Altman, Dario Amodei, Meta’s Alexandr Wang, Sir Demis Hassabis, Pushmeet Kohli, Y Combinator’s Paul Graham, and Mustafa Suleyman. DeepMind explicitly warned that traditional commercial DNA synthesis screening relies on comparing customer orders against databases of known dangerous pathogens—meaning novel, AI-generated protein binders, viral sequences, or toxins can slip straight past legacy screening filters unrecognised.
Whether the paper’s researchers had envisioned the paradigm of artificial intelligence (AI) with life sciences and medicine research, which frequently overlaps, changing rapidly since, or were quietly trying to nudge development in the right direction, the trajectory is clear to see. Zhang, then at Scale AI, has been a Research Program Manager at OpenAI since June. The AI company is part of a troika that includes Anthropic and Google DeepMind, actively researching solutions. Bioscience AI research core focus areas include molecular structures and typical AI tasks include protein folding, genomic sequencing, and metabolic pathway mapping. Medicine AI research focuses on disease diagnosis, clinical trial matching and predicting patient outcomes. AI companies and scientists realise that real safety with AI and biosciences depends on practical checks in the real world, such as verifying synthetic DNA orders, vetting access to specialised AI models and finding dedicated lab efforts. AI is proving beyond doubt that it can speed up the hunt for new medicines and new vaccines. It is upon the industry and governments to ensure biosecurity laws don’t allow the risks to outrun the cures.
“Large language models (LLMs) perform increasingly well on biology benchmarks, but it remains unclear whether they uplift novice users—i.e., enable humans to perform better than with internet-only resources,” Zhang wrote at the time. “Both companies conducted biosecurity-specific evaluations and classified their models at equivalent risk tiers: meaningful uplift for users with some technical background, but not equivalent to expert-level assistance for novel catastrophic weapons development,” said Nikki Teran, PhD and Managing Member, Emerging Technology Solutions LLC, at NTI.
