Anthropic says its biology lab has already found something big

Last week Anthropic confirmed a wet biology lab. This week it says that lab already surfaced what it frames as a major find: a previously unknown enzyme system in bacteriophage DNA with properties reminiscent of CRISPR, able to perform operations like cutting, copying, and pasting DNA.

CEO Dario Amodei stresses the discovery was made mostly, though not entirely, by Claude — and acknowledges a Stanford team previously found a system that is in some ways similar. Independent validation will decide how big or new this actually is.

We’re treating the science claim as provisional and the lab-governance reveal as the harder story.

The speed narrative is striking if it holds: lab opened this spring (Anthropic won’t say how many months), Claude’s concerted run cited at 21 hours, about 950 agents, roughly 210 million tokens. That is a demo of AI-accelerated literature-and-data search feeding human bench work — not a robot scientist. TechCrunch underlines the constraint Anthropic wants on the record: BSL-1 and BSL-2 only, no human-infecting pathogens, all physical experiments performed by human scientists. Amodei has not ruled out Claude eventually controlling equipment with safeguards; they are not doing that today.

The timing sits awkwardly against recent CEO talk about slowing down for safety tests and Amodei’s own stated fear of AI-enabled bioterrorism, alongside his belief AI could cure most diseases in 5–10 years. Anthropic is choosing the upside path with guardrails it describes as human-in-the-loop wet work. Useful AI in biology is exactly accelerated hypothesis search that shortens the path to tools and therapies — CRISPR’s history shows how a bacterial immune trick becomes infrastructure. The harm side is dual-use knowledge, overclaimed novelty, and the slippery slope from “humans run the pipettes” to “autonomous lab equipment” without commensurate oversight.

I’m watching whether outside labs confirm the enzyme system’s novelty and capability relative to the Stanford-adjacent prior art Amodei flagged, and whether Anthropic publishes enough method detail for replication. If the find is real and meaningfully new, Claude-as-research-accelerator gets a concrete win. If it is a thin remix of known phage biology with great marketing, the industry learns the wrong lesson about 21-hour miracles.

Julie Bort for TechCrunch, Sep 23, 2026 (3:17 PM PDT), on Anthropic’s wet-lab enzyme-system claim, Claude’s role and token/agent counts, Amodei’s X acknowledgments, BSL limits, and parallel AI-biology work at Stanford, UCSF, and Google.