Claude has found a previously unknown enzyme system in the DNA of viruses that infect bacteria. Anthropic announced it on 23rd September 2026 as the first result from its new molecular biology lab in the San Francisco Bay Area. The system is called ART, short for array-associated reverse transcriptases. Anthropic doesn’t know what ART does yet.

Earlier this month, we covered Anthropic’s Fable 5.1 and Mythos 5.1 models. The search that found ART ran on the older Claude Mythos 5.

An enzyme is a protein that does one specific job inside a living cell. A reverse transcriptase is an enzyme that copies RNA into DNA. DNA is the genetic code itself, and RNA is a working copy of part of it.

ART lives in bacteriophages, the viruses that attack bacteria, and mostly in very large ones called jumbo phages. It has three parts: the reverse transcriptase, a partner gene right next to it, and a long row of short DNA sequences repeated at even spacing. The enzyme wasn’t new, since earlier studies had already recorded it. Nobody had noticed the repeating row or the partner gene.

CRISPR began the same way. In 1987, researchers in Japan noticed odd repeats in the DNA of E. coli bacteria. In 2012, scientists turned CRISPR into a tool that cuts DNA at a spot they choose, which is what gene editing means. That took 25 years. CRISPR is now the basis of gene-editing medicines.

In a CRISPR system, each gap between the repeats holds a different short sequence, and that bank of sequences lets scientists aim CRISPR at a chosen target. Only a handful of known systems share ART’s mix of features. Every one of them is programmable and can cut, copy or paste DNA. Nobody has shown that ART can do any of that.

Anthropic’s scientists wrote one research brief and handed it to Claude. About 950 Claude agents, each a separate copy of the AI working on its own task, ran for 21 hours and used 210 million tokens, the small chunks of text AI models read and write. They searched 1.9 billion protein clusters built from DNA collected directly from environmental samples. The agents pulled out more than 200,000 reverse transcriptases, flagged 3,500 candidate systems and cut those down to 20 written reports. Anthropic says the same analysis takes an expert weeks to months.

The brief asked for new partner genes and said nothing about repeats. One agent was reading the raw DNA beside an odd-looking enzyme when it wrote that the region was “spectacular”, followed by “that’s a CRISPR-like … repeat array?!” It counted 14 copies of a 16-letter repeat, compared the layout with known systems, searched the published research and filed a report for the human scientists.

Ordinary database software would have missed it, because it only finds the features it’s told to look for. Spotting anything else has depended on a human expert reading the data, and no expert can read billions of sequences.

Human scientists did all the lab work. So far, they’ve shown the ART array is switched on. In published data from a jumbo phage infecting Staphylococcus bacteria, RNA from the array made up as much as 8% of the virus’s RNA. When Anthropic’s team put the system into E. coli, the array split into distinct short RNAs. CRISPR arrays split the same way, which Anthropic says hints that ART could work similarly.

The preprint, a research paper published before other scientists review it, is plain about the limits. The team hasn’t shown the enzyme is active, whether it works with its partner protein, or what the system does for the virus.

Anthropic also ran the same search ten more times, and none of those runs found the repeats. Most of them came across the ART enzyme, but no agent opened the DNA sitting next to it. When Anthropic later gave its newest Claude models that stretch of DNA to read directly, they spotted the repeats in at least nine out of ten attempts.

Feng Zhang, a CRISPR gene-editing pioneer and professor at MIT and the Broad Institute, reviewed the preprint. He called the arrays “genuinely intriguing” and worth further investigation.

There’s no new treatment here and no new tool. What Anthropic has shown is that an AI agent can spot something in raw DNA it wasn’t asked to look for, then check it the way a researcher would, in under a day. Whether ART is useful depends on experiments Anthropic says are underway. It hasn’t said when it will publish results.