The Screening Duty for Synthetic Genomes Ends Where Federal Money Ends
In February 1975 about a hundred and forty molecular biologists met at a conference center on the California coast and agreed to constrain their own work. The Asilomar meeting produced containment categories for recombinant DNA research, written by the people who wanted to do the research. Those categories acquired force sixteen months later. The National Institutes of Health issued them as guidelines on June 23, 1976, published in the Federal Register on July 7, and attached them to its own grants. They were a funding condition. Bills to extend them to companies taking no federal money, among them H.R. 6158 in the 95th Congress, produced no statute.
That shape has held. On August 6, 2026, Science published Generative design of bacteriophages with genome language models, in which a team from Stanford and the Arc Institute reports building working viruses from genomes an AI model wrote. Rules of several kinds can reach a laboratory doing this work: institutional biosafety committee review, the funding conditions that descend from Asilomar, the select-agent lists, export controls. What none of them supplies is a generally applicable federal screening duty on the company that synthesizes the DNA when the buyer is a privately funded domestic customer. That counter is where a designed genome becomes a physical virus.
What the team built
The design template was ΦX174, a lytic bacteriophage of roughly 5,400 base pairs that infects Escherichia coli and has been a laboratory workhorse for fifty years. The authors, led by Samuel H. King with Brian Hie among the senior authors, fine-tuned the genome language models Evo 1 and Evo 2 on 14,466 sequences from Microviridae, the viral family ΦX174 belongs to, then asked for whole genomes with realistic genetic architecture and the same narrow host range.
From the models' candidate pool the team synthesized and tested 285 designs. Sixteen produced viable phages. Several outcompeted ΦX174 in growth competition and lysed their host faster. Cryo-electron microscopy of one design found a DNA packaging protein evolutionarily distant from anything in its training family. A cocktail of the generated phages cleared ΦX174-resistant E. coli across three strains, and that last result is where the medicine sits. Resistance to a given phage is one of several obstacles that have kept phage therapy at the margins of clinical practice, and a method that writes fresh candidates on demand could in time shorten the search for the next one. This is a laboratory proof of concept. No patient benefit has been shown.
Which rules can attach
The team's precautions were substantive. Human-infecting viral sequences were excluded from the training data. The target was restricted to E. coli C, a non-pathogenic laboratory strain. The wet-lab work ran under protocols the Arc Institute describes as exceeding standard requirements, in dedicated biosafety cabinets with specialized disposal. Rules beyond the team's own can attach to work like this, and where they do they are not optional: institutional biosafety committee review under the NIH guidelines, and the purchasing terms that come with federal money. What a given project's funding and approval record contains is a matter for that record, and this analysis does not assert it.
They were also constraints on the researchers. In the same issue of Science, Thomas Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security published a commentary on the paper that puts the gap on the other side of the transaction: the ability to compose viral genomes with generative AI now exists, and the governance able to steer it safely does not. Their proposal is specific. Providers of synthetic nucleic acids should be legally obliged to screen the sequences they are asked to make and the customers doing the asking, in place of today's largely voluntary practice. They add that screening methods capable of flagging genuinely new designs still need to be developed.
Practical takeaway. The control point for AI-designed genomes is the synthesis order counter, and in the United States it is currently reached through a term attached to federal research grants. No generally applicable federal screening mandate binds the companies that fill the orders.
The governing instrument is a procurement condition
The United States does have an instrument here. The White House Office of Science and Technology Policy released its Framework for Nucleic Acid Synthesis Screening on April 29, 2024, effective a year later on April 29, 2025. It requires federally funded life-science researchers to buy synthetic nucleic acids, and the benchtop machines that make them, only from providers who screen. Its legal form is the Asilomar form, half a century on: a condition of receiving federal research money. A buyer outside that funding stream is outside the framework.
The framework remains the current one, and the replacement it was ordered to receive is overdue. Executive Order 14292, "Improving the Safety and Security of Biological Research," issued May 5, 2025, directed federal departments and agencies to revise or replace it. The same order, at section 5, gave agencies 180 days to produce a strategy for research that federal money does not fund, one that was to include comprehensive nucleic-acid synthesis screening. That is the provision aimed straight at the gap in this story, and it is the one still outstanding. The administration has moved on a neighboring file, issuing a policy for stopping high-risk life sciences research on July 20, 2026; the synthesis screening framework is a different document and has not been replaced. The Administration for Strategic Preparedness and Response still carries the revision notice on the framework's own page. Fifteen months.
Congress has two answers in front of it and has enacted neither. S. 3741, the Biosecurity Modernization and Innovation Act of 2026, introduced by Senator Tom Cotton with Senator Amy Klobuchar on January 29, 2026, would direct the Commerce Department to write mandatory screening regulations that bind providers themselves; it has been with the Senate Commerce Committee since. H.R. 3029, the Nucleic Acid Standards for Biosecurity Act, takes the softer route of standards development and has been ordered reported by voice vote, though the Congress.gov status still shows it at introduction. Neither text has passed its chamber.
Eleven months in plain sight
The chronology is the part worth keeping. The manuscript went up on bioRxiv in September 2025 and appeared in Science on August 6, 2026: eleven months between preprint and journal, on a document anyone could read, including the people drafting the replacement framework. That window was the natural moment for a rule to arrive. Nothing arrived in it.
A second problem outlasts the calendar. Order screening rests substantially on comparing a requested sequence against curated databases of known agents of concern, which is a test of resemblance. The paper's central scientific claim is substantial evolutionary novelty: the packaging protein that cryo-EM found sits far from its family. Nothing in this experiment tested or evaded a provider's screening, and it should not be read as having done so. What it does establish is that designed sequences can be genuinely unlike the known ones, which is a reason the comparison-based methods now in use need supplementation. That is precisely what Inglesby and Hanke ask for. Writing a screening duty into law in 2026 is a harder technical assignment than it was when the framework was drafted in 2024, and Evo 2 is public.
How Quentir Reads It
Set this beside what happened in cryptography late last month. A model found a mathematical flaw in HAWK, a post-quantum signature scheme, after roughly two years of human review had passed it, and the designers pulled it out of NIST's additional-signatures round. We covered that withdrawal when it happened. The machine result was startling and the institutional response was dull, which is the compliment: there was a venue, the venue had one defined action, and someone took it within days.
Genome design has the chokepoint without the venue. Synthesis runs through a small number of commercial providers, which is exactly why screening at the order counter is the sensible control. What is missing is a duty on the provider that survives a customer paying its own way, and a body positioned to decide that a particular design does not get made. The Stanford and Arc team brought that judgment with them and, by every published account, brought it well. The arrangement does not require the next buyer to.
This is the pattern we have been tracking all summer across quantum and AI policy, where the instruments that arrive first are procurement conditions, tenders and committees reaching the funded party while the underlying activity stays untouched. Quentir's Signature Brief editions follow that class of instrument one file at a time, with fixed scope, the primary texts, an executive synthesis and the triggers that would change the reading; the current editions are listed on /products, and the wider public archive stays open on the Quentir blog.
Two things are outstanding, and they work differently. The replacement framework runs on federal procurement: it encourages providers and enforces through funding, which leaves the privately funded buyer roughly where it found them. Section 5 of the order points somewhere else, at a strategy for the settings federal money does not fund and at a legislative proposal where the authority to reach them is missing. That last clause is the admission. Putting a screening duty on providers themselves takes a statute, or a regulation issued under one, and S. 3741 is the live candidate for it. Which of these arrives first will answer a question open since 1976: whether the screening of synthetic genomes stays a term of a research grant or becomes a condition of doing business.
Published intelligence, built to inform your own decisions. Published: August 7, 2026.
Sources: Samuel H. King et al., "Generative design of bacteriophages with genome language models", Science (August 6, 2026), DOI 10.1126/science.aec2657; preprint bioRxiv 10.1101/2025.09.12.675911 (September 2025). Thomas V. Inglesby and Moritz S. Hanke, "AI-designed viral genomes", Science (August 6, 2026), DOI 10.1126/science.aej8512. Arc Institute, "How We Built the First AI-Generated Genomes". White House Office of Science and Technology Policy, Framework for Nucleic Acid Synthesis Screening (released April 29, 2024; effective April 29, 2025), page consulted August 7, 2026; Executive Order 14292, "Improving the Safety and Security of Biological Research" (May 5, 2025), including its section 5 direction to produce a strategy for non-federally funded research covering comprehensive nucleic acid synthesis screening; ASPR, United States Government Policy for Stopping High-Risk Life Sciences Research (July 20, 2026). S. 3741, Biosecurity Modernization and Innovation Act of 2026 (introduced January 29, 2026) and H.R. 3029, Nucleic Acid Standards for Biosecurity Act, 119th Congress. National Institutes of Health, Guidelines for Research Involving Recombinant or Synthetic Nucleic Acid Molecules, issued June 23, 1976 and published in the Federal Register on July 7, 1976 following the February 1975 Asilomar conference, per the NIH Guide notice of 1976; H.R. 6158, 95th Congress, as an example of the recombinant DNA bills that did not become law, and NIH guidance on institutional biosafety committee administration.
Published intelligence, built to inform your own decisions. Published: August 7, 2026.