The AI Labs Found a Shared Nightmare

Sam Altman (OpenAI) and Dario Amodei (Anthropic) run their companies like they're locked in a fight to the death for the AI market, and in public they agree on almost nothing.

This week, they signed the same letter, alongside Google DeepMind's Demis Hassabis and Microsoft AI's Mustafa Suleyman: "In Support of Mandatory Nucleic Acid Synthesis Screening and Recordkeeping." The signers run the full AI-policy spectrum. Some warn the technology could end the world. Others call that fear overblown. They rarely line up behind anything, but this was the exception.

They're all pointing at DNA, the blueprint of life. The letter, published June 3, asks Congress to make mandatory what is now voluntary. Companies would have to screen every synthetic DNA and RNA order against databases of dangerous sequences, and keep records so an order that later proves dangerous can be traced to whoever placed it. The biggest, most responsible providers already do both. Smaller and offshore sellers often don't. The hole is everyone else.

The fear is that AI has gotten unnervingly good at the hands-on biology that used to take years of lab training to learn. On a 2025 benchmark of practical virology lab problems, expert virologists averaged 22.1% on questions within their own specialties, while OpenAI's o3 hit 43.8% and outperformed 94% of those experts.

A Microsoft-led study published in Science last October took this further. Researchers used AI protein-design tools to "paraphrase" known toxins like ricin, scrambling the sequence while preserving the structure and active sites, and most of the redesigns could slip past the screening software DNA synthesis companies use. Before they published, the team had already built a fix for that software and pushed it to providers worldwide.

Those same tools are the engine of new medicine. RFdiffusion, one of the models behind work like this, came out of David Baker's lab and underpins Xaira Therapeutics, a drug company he co-founded that launched in 2024 with a billion dollars in funding. Its models design antibodies from scratch that hit targets older methods couldn't, from a protein on the flu virus's surface to a marker behind a childhood cancer.

The threat side is murkier than the medicine. Both studies are proxies, not proof. The virology test is multiple choice, and the Microsoft toxins were never synthesized, so no one has taken an AI design all the way to a working pathogen. But the capability itself isn't in doubt, because the same kind of design is already making medicine. A tool reliable enough to build a therapy is reliable enough to build a threat.

The people making the case also stand to gain from it. The Science study came from Microsoft, whose AI chief signed the letter, and from Twist Bioscience and the industry's own screening consortium. Twist has long pushed to get ahead of this kind of misuse, and it backs the Cotton-Klobuchar bill the campaign wants passed. A mandate rewards the companies already screening, and the vendors who sell them the software, while the cost lands on smaller, thinner-margin rivals. None of that is hidden, and it goes both ways.

That bill, introduced in January, relies on homology-based screening. It matches each order against a list of known dangerous sequences. But the danger the letter keeps pointing to has no match on any list. AI now designs proteins from scratch, unlike anything in nature, and that isn't a flaw in the tools. It's the whole point. The very thing that lets Baker's models reach undruggable targets is what makes them invisible to a list-based screen. The strength is the loophole.

David Baker won the 2024 Nobel Prize in Chemistry for that work, and he signed this letter too. He and Harvard's George Church had flagged this exact gap years ago and proposed two defenses. Log every order, so a new threat can be traced after it ships. And screen for what a protein does, not just what it resembles, so a dangerous one is caught before it ships.

The bill takes the logging and leaves the rest. The screening that would actually catch a novel design is handed to a federal agency to research, with no deadline and no requirement to use what it finds. Meanwhile, the labs designing these proteins keep getting faster, with real money behind them. Only one side of this has a billion dollars.

Select the right answer! (See explanation below and source)

Mount Sinai scientists found a hidden drug-binding pocket on a cancer protein that leading AI structure tools failed to predict. Published in JACS, the study identified the site on PKMYT1, a kinase tied to cell division, and confirmed it with X-ray crystallography after AlphaFold2, AlphaFold3, and Boltz-2 all missed it. [Link]

The FDA just accepted an AI tool for predicting drug-related liver damage into a formal qualification pathway. Its Center for Drug Evaluation and Research admitted the AI-driven liver model to the ISTAND pilot, the first step in a multi-stage process that could eventually let drugmakers cite the tool in regulatory submissions. [Link]

Novo Nordisk and BridgeBio are already using a newly launched AI platform that turns a large human genetics database into a tool for finding drug targets. UK company Genomics built the platform, called Mystra AI, around a plain-language chat interface so scientists can query genotype-phenotype data for target discovery and validation without specialist analytical skills. Novo Nordisk's VP for AI and digital innovation in R&D said the drugmaker turned to Mystra after concluding it couldn't build a comparable platform in-house. [Link]

iBio just dosed the first patient in a human trial of IBIO-600, the lead asset from its AI-driven drug discovery platform. The milestone marks iBio's transition to a clinical-stage biotechnology company, and its CEO credited the AI-integrated platform with moving the program from initiation to the clinic in about two years. IBIO-600 is a long-acting antibody designed to preserve muscle during weight loss and to be dosed only a few times a year, including potentially alongside GLP-1 drugs. [Link]

Because the two DNA strands run in opposite directions, one strand can be copied continuously, while the other must be copied in pieces called Okazaki fragments. [Source]