AI CEOs ask Congress to make DNA synthesis screening mandatory

What happened

On June 5, 2026, the CEOs of four of the largest AI developers — Dario Amodei (Anthropic), Sam Altman (OpenAI), Demis Hassabis (Google DeepMind) and Mustafa Suleyman (Microsoft AI) — co-signed a public letter to the US Congress, alongside dozens of life-sciences and national-security experts, asking lawmakers to make synthesis screening legally mandatory for sellers of synthetic DNA and RNA. The letter, hosted at screendna.org, was organized by the Foundation for American Innovation and the Institute for Progress. Notably, several nucleic-acid synthesis companies — including Twist Bioscience and Ansa Biotechnologies — also signed, indicating that part of the industry welcomes the regulation it would be subject to.

The ask is narrow and concrete: require every provider that synthesizes DNA or RNA to screen each order against databases of known hazardous sequences, verify the identity of the customer placing it, and keep records of orders and the exact specifications of what was sold. Voluntary screening already exists, but coverage is uneven; the letter argues that uneven, opt-in controls are not a credible defense if the underlying risk is rising.

Why AI changes the biosecurity calculus

The core argument is not that AI invents new pathogens, but that it lowers a knowledge barrier that historically did much of the work of keeping biological weapons rare. The signatories write that “the knowledge barriers which have historically prevented bad actors from obtaining biological weapons” may “meaningfully erode” as models improve. That claim has independent support: a New York Times report on April 29, 2026 documented that publicly available AI models could supply information relevant to building and spreading biological weapons, and frontier-model evaluations have repeatedly shown models answering expert-level wet-lab questions.

Scale compounds the concern. Stanford HAI’s 2026 AI Index found that generative AI tools reached an estimated 53% of the world’s population within three years — faster diffusion than either the PC or the internet. A capability that is both more useful and far more widely distributed shifts the threat model from “a handful of state programs” toward “a much larger population with partial uplift,” which is precisely the regime where a physical chokepoint matters most.

The synthesis chokepoint

Designing a sequence is information; turning it into physical material requires synthesis. That makes the synthesis step a natural control point — analogous to how the security of an LLM agent often comes down to controlling which actions untrusted input can trigger, not to perfectly filtering the input itself. If every order must be screened against hazard databases and tied to a verified customer, the pipeline from “a model described something dangerous” to “someone has the physical agent” gains a gate that does not depend on the model’s own guardrails holding.

This is the same defense-in-depth logic that security teams apply to AI systems: assume upstream controls (here, model refusals and alignment) will sometimes fail, and place an independent, harder-to-bypass control downstream at the point where harm becomes physical.

Defenses and what this means for builders

For policymakers and the synthesis industry, the letter’s concrete asks map onto a defensive checklist: mandatory sequence-of-concern screening on every order; customer verification (know-your-customer) so orders are attributable; record-keeping of orders and specifications to support later biosecurity investigations; and statutory exemptions for clearly non-hazardous materials so legitimate research is not throttled.

For teams building on LLMs, the takeaway is architectural rather than biological. First, treat model-level safety as one layer, not the layer: refusal training and red-teaming reduce uplift but should never be the only barrier between a capability and real-world harm. Second, where your product touches a regulated physical or financial chokepoint, integrate with the screening and KYC controls that exist there rather than reinventing them. Third, log and retain enough provenance to support an investigation — the same record-keeping the letter asks of synthesis providers is good practice for any agentic system whose actions have external consequences. None of this requires, or benefits from, any operational detail about the agents themselves; the defensive value lives entirely in the process controls.

Status

The letter supports, but is separate from, pending legislation. The Biosecurity Modernization and Innovation Act of 2026 (S.3741), introduced in February 2026 by Senators Tom Cotton (R-Ark.) and Amy Klobuchar (D-Minn.), would require sellers of synthetic nucleic acids to screen both orders and customers, with exemptions for materials judged clearly non-hazardous. As of mid-June 2026 the bill remains in the legislative process; the June 5 letter is widely read as an effort to demonstrate cross-industry alignment — AI developers and synthesis firms together — to help move it forward. Existing US law (the Biological Weapons Anti-Terrorism Act of 1989, expanded after the 2001 anthrax attacks) already criminalizes developing or possessing biological agents as weapons, but does not mandate synthesis screening — the gap this proposal targets.

This article covers a public-policy and biosecurity-governance development. It deliberately contains no operational detail about biological agents. Sources are cited with their publication dates above.