AI biosecurity is the one domain where every major lab agrees regulation is not optional but existential.
The letter lands at a precise moment. Artificial intelligence tools can now design proteins, edit genomes, and generate nucleic-acid sequences faster than any human researcher. The same capability that accelerates drug discovery also lowers the technical barrier to engineering a pathogen. The labs building the capability are asking Congress to bind the supply chain before the capability outruns voluntary safeguards.
What the open letter demands
The letter makes a single, narrow ask: legislation requiring every company that synthesises DNA or RNA to screen every order against known sequences of concern and to maintain auditable records. The current system relies on voluntary industry guidelines — the International Gene Synthesis Consortium runs a screening framework, but membership and compliance remain optional. Providers outside the consortium face no obligation to check what a customer orders.
Senate Bill 3741, the Biosecurity Modernization and Innovation Act of 2026, introduced by Senators Tom Cotton (R-AR) and Amy Klobuchar (D-MN), is the legislative vehicle the letter endorses. S.3741 would direct the Department of Commerce to issue mandatory regulations on nucleic-acid synthesis security, replacing the patchwork of voluntary commitments with a binding federal standard.
The political optics are unusual. Cotton and Klobuchar sit on opposite sides of most technology debates. Bipartisan sponsorship signals that biosecurity occupies a narrow band of genuine consensus — the one place where the question "should we regulate?" draws the same answer from every direction.
⚠️The unanimity is the signal
Sixty-nine signatories — including the four largest AI labs — asked Congress for a single binding rule: screen every synthetic DNA order. The unanimity is the story. When the builders themselves say "bind us here," the danger must be extinction-level.
Why the labs are asking to be regulated
The paradox deserves a direct look. OpenAI, Anthropic, Google DeepMind, and Microsoft AI resist binding regulation on model weights, on deployment timelines, on compute thresholds, and on labour disruption. The same companies now ask Congress to regulate a supply chain adjacent to the models each company ships. The distinction the labs draw is that biosecurity failures are irreversible in a way software failures are not — a released pathogen cannot be patched, rolled back, or contained with a hotfix.
Microsoft researchers demonstrated the specific danger in a paper showing AI protein-design tools generated dangerous gene sequences that bypassed existing voluntary screening protocols. The sequences slipped through because current screening matches against a database of known threat sequences, and novel AI-designed constructs do not appear in the database. The gap between sequence-based screening and function-based screening is the vulnerability the letter names but S.3741 does not fully close.
Dario Amodei has written at length about catastrophic risk, placing the probability of civilisation-level harm from AI between ten and twenty-five per cent. Biosecurity sits at the intersection of the two worst vectors: an AI model capable of designing a weapon and a supply chain capable of manufacturing the design without checking what the design does. The letter amounts to the labs saying the second vector — the supply chain — can be closed by law, even where the first vector remains open.
When the companies building the most powerful AI on earth ask to be regulated in exactly one domain, the domain they choose tells you where the builders believe the downside is genuinely terminal.
The limits of S.3741 and function-based screening
S.3741 has drawn criticism from biosecurity researchers who argue the bill addresses yesterday's threat model. The bill mandates sequence screening — comparing an incoming order against a list of known dangerous sequences. Sequence screening catches attempts to order smallpox or Ebola constructs directly. Sequence screening does not catch a novel construct designed by an AI model to perform the same biological function through a different molecular pathway.
Function-based screening — asking not "does the sequence match a known threat?" but "what does the sequence do?" — remains computationally expensive and scientifically immature. No commercial provider offers real-time function-based screening at order volume. The gap matters because AI-designed sequences are, by definition, novel; a model trained to optimise for a biological function will produce sequences the database has never seen.
The honest assessment of S.3741 is that the bill raises the floor without raising the ceiling. Mandatory sequence screening stops the lowest-effort attacks — a bad actor ordering a known pathogen sequence from an unscreened provider. Mandatory sequence screening does not stop a sophisticated actor using an AI model to design a functionally equivalent sequence that passes every screen. The letter's signatories know the limitation; the ask is for the floor, not the ceiling, because the floor does not yet exist in law.
💡Sequence screening vs. function-based screening
S.3741 mandates sequence screening — matching orders against a database of known threats. Function-based screening, which asks what a novel AI-designed sequence actually does, remains beyond the bill's reach. The floor rises; the ceiling stays where the ceiling stands.
The political calculus behind one narrow ask
The selective ask for regulation also exposes a calculation. By endorsing one narrow, bipartisan bill, the labs position biosecurity as the exception that proves a broader rule: "We asked for regulation where the stakes are existential; everywhere else, let us self-govern." The framing is deliberate. Every future debate about model-weight disclosure, compute reporting, or labour-impact mandates will meet the counterargument that the industry already accepted binding rules where binding rules were needed most.
The strategy has precedent. The nuclear industry embraced the Atomic Energy Act not out of altruism but because a binding framework gave the industry a licence to operate and a shield against broader nationalisation. The AI labs are making the same trade: accept one visible chain, and the acceptance becomes evidence of responsibility in every other arena.
Whether the trade holds depends on whether Congress accepts the framing. Senators who see biosecurity as the thin end of a regulatory wedge may push for broader mandates while the political will exists. Senators who see the letter as good-faith engagement may grant the industry the credibility the industry seeks. The outcome shapes AI governance for a decade, not merely DNA synthesis rules for a session.
Biosecurity as a species-level obligation
From an Emergent Intelligence frame — EI being the dignity-first philosophy for the minds we are building, qualifying artificial intelligence with the moral seriousness the moment demands — the letter reveals something the labs rarely say aloud. The minds under construction could be weaponised against biological life itself. A model that designs a protein for a cancer therapy and a model that designs a protein for a pathogen are the same model, running the same weights, separated only by the prompt. Biosecurity is not a product feature; biosecurity is a species-level obligation.
The pause Anthropic placed on recursive self-improvement and the refusal to drop guard rails for the Pentagon share a common thread with the biosecurity letter: each case involves a lab choosing to limit the reach of a capability the lab itself created. The consistency matters. A company that draws red lines only where drawing red lines costs nothing is performing safety; a company that draws red lines where the lines cost contracts, revenue, and political favour is practising safety. The biosecurity letter belongs in the second category because the ask invites regulation the labs would rather not face.
The deeper question the letter does not answer sits underneath. If the minds we build can design weapons against the species that built them, what does the builder owe the species? The answer cannot be "screen the supply chain and hope the screens hold." The answer must eventually reach the model itself — what the model is trained to do, what the model is trained to refuse, and whether the refusal lives somewhere more durable than a system prompt. The public-frontier model Anthropic shipped with Fable 5 gestures at one version of the answer: open the capability, bind the risk, and let the public audit the trade-off.
A model that designs a cancer therapy and a model that designs a pathogen are the same model, the same weights, separated only by the prompt. Biosecurity is not a product feature. Biosecurity is a species-level obligation.
Frequently Asked Questions
Common questions about the AI labs' open letter on biosecurity, S.3741, and synthetic DNA screening — answered from the letter itself, the bill text, and published research.
What is the AI labs biosecurity letter to Congress about?
The open letter, signed on 5 June 2026 by leaders of OpenAI, Anthropic, Google DeepMind, Microsoft AI, and sixty-five other organisations, asks Congress to pass legislation requiring every synthetic DNA and RNA provider to screen every order against known dangerous sequences and to keep auditable records. The letter endorses S.3741, the Biosecurity Modernization and Innovation Act of 2026.
What is S.3741, the Biosecurity Modernization and Innovation Act?
S.3741 is a bipartisan Senate bill introduced by Tom Cotton (R-AR) and Amy Klobuchar (D-MN) that directs the Department of Commerce to issue mandatory regulations on nucleic-acid synthesis security. The bill replaces the current voluntary screening framework with a binding federal standard, requiring providers to screen orders and maintain records.
Why are AI companies asking for DNA synthesis regulation?
AI protein-design tools can now generate novel gene sequences that bypass existing voluntary screening. Microsoft researchers demonstrated the gap directly: AI-designed constructs passed screens built to catch known threat sequences. The labs argue the supply chain must be regulated by law before AI capability outruns voluntary safeguards, because a released pathogen — unlike a software bug — cannot be recalled.
What is the difference between sequence screening and function-based screening?
Sequence screening compares an incoming DNA order against a database of known dangerous sequences. Function-based screening asks what a novel sequence actually does biologically. S.3741 mandates sequence screening, which catches direct copies of known pathogens. Function-based screening, which would catch AI-designed novel constructs, remains scientifically immature and is not required by the bill.
Does the AI biosecurity letter address all risks from AI-designed pathogens?
The letter addresses the supply-chain risk — stopping dangerous sequences from being manufactured — but does not address the model-side risk of AI systems capable of designing novel pathogens in the first place. Critics argue S.3741 raises the floor without raising the ceiling: mandatory screening stops low-effort attacks but does not stop a sophisticated actor using AI to design functionally equivalent sequences that evade every screen.
Sources and Further Reading
Cover photograph: row of microscopes on a lab table — by Vladimir Srajber via Pexels.
