Anthropic Says Claude Helped Discover a Novel Enzyme System

Claude Helped Researchers Break Into OpenAI Staff ChatGPT Accounts in Under 72 Hours

Anthropic said on Wednesday, September 23, 2026, that its AI model Claude helped find a previously undescribed enzyme system in the DNA of bacteriophages — viruses that infect bacteria. The company compared parts of the find to patterns seen in CRISPR, the gene-editing tool.

It is the first public result from Anthropic’s new life-sciences group and wet lab in the San Francisco Bay Area.

Advertisement

What Claude found

Researchers asked Claude to search huge DNA databases for new examples of reverse transcriptases (RTs) — enzymes that copy RNA into DNA.

About 950 Claude agents ran for roughly 21 hours and used more than 210 million tokens. One agent flagged a repeating stretch of DNA next to an unusual RT gene. Humans then reviewed the lead and ran lab tests.

Anthropic named the system array-associated reverse transcriptases, or ART. It has three parts:

Advertisement
  • the reverse transcriptase
  • a nearby partner protein whose job is still unknown
  • a long array of evenly spaced DNA repeats that look a bit like a CRISPR array

The underlying RT itself had already shown up in earlier studies. Anthropic says Claude was first to notice the wider system: the repeat array and the extra protein.

What it does not mean yet

Anthropic does not yet know what ART does. The company says only a few known systems share this kind of layout, and those systems can cut, copy, or paste DNA. That is a hint, not proof that ART is a new CRISPR.

The write-up is an early company report, not a finished, peer-reviewed paper. Outside scientists will need to check how new the system is and whether it can be used as a tool.

Advertisement

CEO Dario Amodei has called the result preliminary and said it might one day matter for gene therapy — if later work confirms a useful mechanism.

Why Anthropic is highlighting it

The company wants to show that AI agents can scan data at a scale that would take a human expert weeks or months, then hand the best leads to a lab. Scientists still wrote the first prompt and did the wet-lab work.

Source: Anthropic, Reuters, The Verge, TechCrunch

Advertisement