Anthropic says Claude agents found a new CRISPR-like enzyme system in phage DNA

Anthropic says its Claude models have surfaced a previously uncharacterized enzyme system in bacteriophage DNA. Bacteriophages are viruses that infect bacteria. The company describes the system as having properties reminiscent of CRISPR, and its researchers say it can cut, copy and paste DNA. It is the first announced result from the wet biology lab Anthropic confirmed it was running only last week.

The company frames the finding as mainly the work of AI. Anthropic CEO Dario Amodei wrote on X that it came "mostly, though not entirely, by Claude," TechCrunch reported.

How the search worked

According to Anthropic, human involvement in the discovery phase was limited to writing the initial prompt and later doing the lab work. The company says about 950 Claude agents combed a large database of DNA sequences for 21 hours, consuming roughly 210 million tokens. One agent noticed an odd repeating pattern and flagged it for a human to review. Follow-up analysis and experiments then identified the enzyme system, The Verge reported.

Anthropic concedes it does not yet know what the system actually does. It says it chose to publish early to show off Claude's abilities and to give other researchers a view of its work.

Claims still need outside checks

The novelty of the result is not settled. Amodei acknowledged that the work built on earlier research and that a Stanford team had previously described a system with some similarities to the one Claude found. Whether the new enzymes have any practical use, let alone one on the scale of CRISPR gene editing, is unknown, and validation will fall to the wider research community.

Anthropic is also far from alone in applying AI to biology. Stanford researchers recently published work pairing large language models with CRISPR, a UC San Francisco group has used AI to design enzymes from scratch, and Google's AlphaFold arrived in 2020.

A lab with humans at the bench

The lab itself is described as an ordinary molecular biology setup in the Bay Area, founded this spring, though Anthropic would not say exactly when. The company says it works only at biosafety levels 1 and 2, handles no pathogens that can infect people, and relies on human scientists for every physical experiment.

That design choice matters because of the wider debate over risk. Amodei has named AI-enabled bioterrorism as one of his biggest fears, even as he predicts AI could cure most diseases within five to ten years. He has not ruled out Claude one day running lab equipment itself, with safeguards, but says Anthropic is not doing so now.

The timing is strategic as well. The Verge noted that the announcement comes as Anthropic prepares to go public, tries to recruit more scientists, and eyes areas such as drug discovery, part of a broader push by AI labs to prove their models' worth through science.