Overview
- A study published this week in Science by Stanford and the Arc Institute reports that genomic language models called Evo 1 and Evo 2 produced thousands of candidate genomes, 285 were synthesized and 16 of those synthetic phages replicated in Escherichia coli.
- One AI‑designed candidate named Evo‑Φ69 outcompeted the natural reference phage ΦX174 in lab competitions by replicating up to 65 times its starting value and a cocktail of synthetic phages overcame resistance in multiple E. coli strains.
- The authors released the models, code and weights with an explicit warning that the same methods could pose dual‑use risks if applied to human‑pathogen datasets, because retraining or reuse could enable harmful designs.
- Scientists and policy experts responded by calling for concrete safeguards such as mandatory DNA‑order screening, risk‑tiered access controls for powerful biodesign models, traceable genetic watermarks and stronger international verification mechanisms.
- Researchers stress limits to near‑term misuse because the synthetic phages were 90–99% similar to the template ΦX174 and because creating human pathogens would require different datasets, advanced labs and specialist expertise, but they warn these defenses may erode as data and models improve.