AI Designed 285 Synthetic Virus Genomes. 16 Worked. Johns Hopkins Says No One's Governing What Just Happened. Here's Everything Else That Shipped.

Stanford and Arc Institute researchers had genome language models Evo 1 and Evo 2 design 285 synthetic bacteriophage genomes end to end; 16 came back as functional, working viruses — a real result, reported everywhere as "the first fully AI-generated virus," and complicated by its own numbers and a same-issue Science commentary saying nobody currently governs that capability. Full breakdown here. Also this week: a phage-evolved botulinum-toxin protease that triggers pyroptosis in cancer cells and slows tumor growth in a treatment-resistant mouse model, and a pathology foundation model that matches clinical-grade diagnostic tools without further training.

Biotech: an AI-designed virus, and the biosecurity commentary published next to it

"Generative design of bacteriophages with genome language models" (King, Driscoll, Li, Guo, Merchant, Brixi, Wilkinson, Hie, Science, DOI 10.1126/science.aec2657, August 6) used Evo 1 and Evo 2 to generate roughly 700,000 candidate genomes, synthesized 285 of them, and confirmed 16 as functional bacteriophages that infected and killed E. coli, some outperforming the natural template and a 16-phage cocktail overcoming two resistant strains. Thirteen of the 16 carried mutations absent from any known natural sequence. In the same issue, a commentary from Johns Hopkins Center for Health Security researchers Thomas Inglesby and Moritz Hanke argued the legal screening regime for this exact capability doesn't exist: no US requirement yet forces DNA-synthesis providers to screen for AI-generated sequences, and gain-of-function policy covers wet-lab work on natural pathogens, not computational design. We go deep on the hit rate, the novelty question, and the governance gap in this week's feature.

Biotech: directed evolution turns a neurotoxin into a cancer therapeutic

"Evolution of botulinum neurotoxin serotype X proteases to induce inflammatory cell death in cancer cells" (McCreary, Hemez, Raymond, Blum, Gonzalez-Valero, Augustin et al., David R. Liu lab, Nature Biotechnology, DOI 10.1038/s41587-026-03243-9, published online August 3) used phage-assisted evolution to reprogram a botulinum neurotoxin serotype X protease to cleave procaspase-1 and gasdermin D instead of its natural substrate, triggering pyroptosis — inflammatory cell death — in cancer cells. The evolved protease drove roughly 86% lytic cell death and over 90% caspase-1 activation in target cell lines, while remaining only 5-fold less efficient than the wild-type toxin on its natural substrate. Reconstituted with BoNT's native self-delivery domain, it selectively killed cancer cells while sparing non-cancerous ones, and slowed tumor growth (P < 0.0001) in a KPCY pancreatic mouse model that resists combination immunotherapy.

MedTech & HealthTech: a pathology model that matches clinical-grade tools cold

"End-to-end multimodal pathology foundation model with clinical dialogue" (Vorontsov, Shaikovski, Casson, Viret, Zimmermann et al., Paige and Microsoft Research, Nature Medicine, DOI 10.1038/s41591-026-04521-4, published online July 31) introduces PRISM2, trained on 2.3 million whole-slide images and 14 million diagnostic question-answer pairs drawn from roughly 700,000 real pathology reports. With prompt-based inference and no further training, PRISM2 matches or exceeds the balanced accuracy of Paige's own clinical-grade prostate- and breast-cancer detection products and outperforms Paige BLN, while also handling biomarker and outcome-prediction tasks through the same clinical-dialogue interface. The model weights are open on Hugging Face.

What this means for reproducible, local-first science

The throughline this week isn't capability — it's what happens after a capability claim ships. A directed-evolution therapeutic and a pathology foundation model both come with the receipts attached: cell lines, mouse models, held-out accuracy numbers, open weights. The AI-designed-virus story is the harder case precisely because the paper did the same thing — it shipped its own governance gap in writing, in the same issue, rather than waiting for outside critics to find it. That's the standard worth holding every capability claim in this beat to: show the denominator, and show what still isn't checked.