On Thursday, Stanford chemical engineer Brian Hie and bioengineering graduate student Samuel King published a paper in Science showing they used an AI model called Evo 2 to generate bacteriophage genomes — complete viral "recipes" — from scratch. Of nearly 300 designs they synthesized and tested in the lab, 16 produced working viruses that killed E. coli bacteria more effectively than the natural phage they used as a starting point — including E. coli strains that had become resistant to it. The viral sequences have "patterns distinct from anything found in nature." Evo 2 is a genome language model trained on DNA sequences across more than 128,000 genomes, developed by Arc Institute with Stanford, NVIDIA, and collaborators. It's open-source. In the same issue of Science, Johns Hopkins biosecurity experts Thomas Inglesby and Moritz Hanke published a commentary warning that the governance to safely steer this technology "does not exist."

1. This Fights the Antibiotic Crisis (Brian Hie, Stanford; Jef Boeke, NYU Langone)

These viruses kill bacteria that antibiotics can no longer touch — and that's the whole point.

Antibiotic resistance kills 1.27 million people a year. The AI-designed phages are built specifically to kill those resistant bacteria. Hie's lab is already working toward phages that target MRSA and Pseudomonas aeruginosa, two of the most deadly drug-resistant hospital infections.

Bacteria can't adapt to 16 distinct phages at once. Hie explained the logic: "If the bacteria gains resistance to a single phage, it's game over. But if you have multiple genetically distinct phages in a mixture, it's harder for bacteria to develop resistance to the entire cocktail." That variety is what makes a cocktail viable — traditional lab methods can't generate 16 structurally different phages at that scale.

The AI didn't run this experiment alone. Hie's team designed and controlled every step of the pipeline. They built Evo 2 without human-pathogen data from the start — specifically to prevent it from generating anything that could infect humans. Jef Boeke of NYU Langone called the result an "impressive first step" toward AI-designed life forms. Samuel King, the graduate student who ran much of the lab work, said: "New doors in science are now open."

2. But the Safety Rules Don't Exist Yet (Thomas Inglesby and Moritz Hanke, Johns Hopkins Center for Health Security)

The law hasn't kept up with what AI can now do in a biology lab.

AI can now compose functional viral genomes from scratch. That's what Inglesby and Hanke wrote in their Science commentary published alongside the research Thursday. Their direct quote: "The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not." They argued it's no longer a question of "whether" this capability exists, but whether it can be deployed without "enabling serious harm."

No U.S. law requires DNA synthesis companies to screen orders. Companies voluntarily check orders against databases of known dangerous sequences — but AI-generated sequences, like Evo 2's designs, have no natural analogs. They won't match anything in those databases. Hanke said there's "just a huge disconnect" on this regulatory gap. Inglesby and Hanke are calling for a legal requirement that DNA synthesis providers verify both the sequences they're printing and the identity of the person placing the order.

The safety filter in Evo 2 can be reversed. The team excluded human-pathogen data from the training set — a real safeguard. But the researchers' own paper notes "a training filter can be undone." A bad actor could retrain a similar model without those restrictions. More than 100 researchers from Hopkins, Oxford, Stanford, Columbia, and NYU co-signed an open letter calling for restricted access to dangerous biosecurity datasets. The framework has no legal force.

3. The Risk Is Smaller Than It Sounds, Though (Tom Ellis, Imperial College London; Hsu Li Yang, Asia Centre for Health Security)

This is the simplest possible genome to design with AI — human pathogens are exponentially harder.

Bacteriophage genomes are the simplest genomes you can design. Tom Ellis at Imperial College London says phage genomes are "literally the smallest and easiest genome to make." The ΦX174 phage used in this study has fewer than 6,000 base pairs. The COVID genome is six times longer than ΦX174 and "exponentially more complex" to engineer, Ellis notes. What Evo 2 did with a minimal 11-gene phage doesn't transfer to human-infecting pathogens in any obvious way.

Designing the genome is the easy part. Hsu Li Yang at the Asia Centre for Health Security in Singapore says "advanced laboratory capabilities remain substantial barriers to weaponization." Synthesizing a working virus, stabilizing it, and deploying it requires specialized expertise well beyond downloading Evo 2. Bad actors who want to modify existing pathogens can already do that without AI — and that route remains more accessible.

Where This Lands

Senators Tom Cotton and Amy Klobuchar introduced S.3741 in January to fix the DNA synthesis screening gap — mandatory sequence screening, customer identity verification, no split orders across providers. It's pending in the Senate. The Trump administration has rolled back Biden-era AI safety requirements in the meantime rather than adding new ones. Inglesby and Hanke say we need law before a more capable model applies this technique to genomes more complex than bacteriophages. Hie's lab says the biology itself draws that line — his viruses only kill bacteria, and antibiotic resistance is killing people right now.

Sources