Claude agents flag a CRISPR-like enzyme system in 21 hours, though Anthropic can’t yet say what it does

Anthropic said on Wednesday that Claude agents flagged an uncharacterized enzyme system in viral DNA after about 21 hours of searching.

The system has a CRISPR-like pattern, but the company says it doesn’t know what the system does yet. It is one of the first results from a molecular biology group Anthropic set up in spring 2026.

Ten reruns of the same search all missed the repeat array

The team has named the find array-associated reverse transcriptases, or ART. Bacteria combat viruses by using reverse transcriptases, which convert RNA into DNA.

Beside it sat what caught Claude’s eye, an evenly spaced run of repeat DNA next to a second gene of unknown purpose.

That three-part layout echoes CRISPR, which stores a library of RNA guides in an array. Each ART array contains 3 to 21 copies of a short repeat and no Cas genes are situated near any ART locus, the preprint states.

Public RNA data from a Staphylococcus phage reveal the array transcribed into distinct short RNAs. They were up to 8% of the phage RNA 15 minutes post-infection.

The preprint notes the team hasn’t demonstrated the enzyme is active or that it acts on those RNAs.

Anthropic gave the agents a brief. Hunt through 1.9 billion protein clusters for new reverse transcriptase systems.

Each task was planned and run by one worker agent and reviewed by a supervisor agent. As the findings accumulated, supervisors opened new tasks.

The campaign ran for 21.5 hours, the preprint says, with no human input, across 949 agent sessions and 215.6 million tokens. Anthropic’s post rounds these to about 950 agents and 210 million tokens.

Agents amassed ~200,000 enzyme clusters and scored 3,564 candidate partner families. The campaign closed with 19 reports for human review.

ART surfaced almost by accident. An agent reading raw DNA beside a peculiar enzyme wrote that it could see a repeat pattern “by eye,” then counted the repeats and looked them up in the literature before writing it up.

Anthropic repeated the same campaign ten more times. All missed the array, no rerun read the DNA upstream of the enzyme.

When given the DNA directly, the best models described the array in a minimum of 90% of attempts, but with the addition of tools and files, the rate sank to as low as 32%.

Feng Zhang backs more study as Kevin Blake doubts any therapy

Feng Zhang, a CRISPR pioneer at MIT and the Broad Institute, reviewed the preprint. He called the identification of RNA-repeat arrays tied to reverse transcriptases “genuinely intriguing,” saying it “merits further investigation.”

On X, CEO Dario Amodei went a step further. The “molecular machine” could “represent a new gene editing mechanism,” he suggested, in keeping with his bet that AI could compress decades of biological progress into years.

ART came out of a Bay Area lab Anthropic only recently made public. The company says the facility works at biosafety levels 1 and 2, does not handle pathogens that infect humans and leaves all bench work to human scientists.

As Cryptopolitan reported when the wet lab was confirmed, it is testing to see if Claude can help direct experiments.

In April, Anthropic purchased startup Coefficient Bio for about $400 million. That same month, OpenAI shipped its own life sciences model, GPT-Rosalind.

Anthropic says experiments to nail down what ART does are already underway. It is inviting other scientists to put forward proposals to conduct research.