Anthropic says Claude found a CRISPR-like enzyme system in a 21-hour search

Anthropic has announced the first significant result from the biology laboratory it confirmed running in the Bay Area: an enzyme system that the company believes behaves like CRISPR, and which it credits largely to its Claude AI models.

A CRISPR-like system inside bacteriophages

According to TechCrunch, the find is a previously undocumented enzyme system embedded in the DNA of bacteriophages — viruses that infect and replicate inside bacteria. Anthropic says the system parallels CRISPR, the bacterial immune mechanism that researchers turned into a widely used gene-editing toolkit, because it can splice, copy and insert stretches of DNA.

The comparison is Anthropic's own characterisation, and it has not been independently verified. As TechCrunch points out, judging how novel or consequential the discovery really is will fall to the broader research community.

Twenty-one hours of machine effort

The lab was only established this spring, and Anthropic declined to tell TechCrunch exactly how many months it has been operating. Even so, the company says the discovery itself required just 21 hours of concentrated effort by Claude, which deployed roughly 950 agents to comb through data and processed around 210 million tokens in the process.

CEO Dario Amodei described the result as mostly, though not entirely, the work of Claude. He also acknowledged that the discovery rests on the contributions of others, writing on X that a Stanford team had previously found a system that resembles the one Claude identified in some respects.

Humans still hold the pipettes

Perhaps the most consequential detail in the announcement is what Claude did not do. Anthropic says the physical experiments were performed entirely by human scientists in a lab that handles only lower levels of biosafety risk, BSL-1 and BSL-2, and does not work with pathogens capable of infecting humans.

Amodei has not ruled out greater autonomy later, saying that Claude may eventually be able to run experiments by controlling lab equipment itself, provided appropriate safeguards exist — but he was clear that is not happening today.

An awkward moment for safety talk

The timing is notable. The existence of the wet lab became public shortly after AI leaders, Amodei among them, publicly conceded that models have grown capable enough, and potentially hazardous enough, that the industry should slow down and build safety-testing procedures. TechCrunch also recalls that some Anthropic employees had voiced fears this month about existential risks from AI.

Amodei has himself named biologically enabled terrorism as one of his gravest concerns, while also predicting that AI could eliminate most diseases within five to ten years. Anthropic, in other words, has judged the potential upsides worth the risk.

AI-assisted biology is also far from unique to Anthropic. TechCrunch notes that Stanford researchers have just published work pairing large language models with CRISPR, a team at UC San Francisco has used AI to design enzymes from scratch, and Google has been in the field since launching AlphaFold in 2020.

Why it matters

This is one of the clearest public demonstrations yet of an AI model driving an autonomous scientific research process end to end — from searching vast biological datasets to reaching a candidate discovery — inside a lab owned by the model's own developer, and within months of that lab opening. If the finding survives independent validation, it strengthens the case that AI systems can accelerate discovery in the life sciences rather than merely assist with it.

At the same time, the caveats matter as much as the headline. The discovery awaits external scrutiny, similar systems were already known, and the hands-on laboratory work remains human. The story is therefore less about an immediate scientific breakthrough and more about a credible proof of concept: AI-driven experimentation is no longer hypothetical, and the debate over how quickly to automate it — and how safely — is now concrete rather than theoretical.