Anthropic’s Claude Discovers a Novel CRISPR-Like Enzyme System

On Sept. 23, 2026, Anthropic announced that its Claude AI had independently discovered a previously uncharacterized enzyme system embedded in the genomes of bacteriophages — the viruses that infect bacteria. The system, which the team named array-associated reverse transcriptases (ART), sits next to a long, evenly spaced array of repeating DNA sequences, a layout that closely recalls the architecture of CRISPR. The finding is the first public result from a new internal molecular biology lab Anthropic quietly assembled in spring 2026.

The scale of the computational search is striking. Roughly 950 Claude agents ran in parallel for 21 hours, burning through 210 million tokens of data while screening more than 200,000 reverse transcriptase sequences. From that pool, the agents flagged 3,500 candidate systems, then narrowed those to 20 compelling cases with human-readable summary reports — work that Anthropic estimates would take an expert scientist weeks to months to complete manually. Anthropic scientists provided only the initial high-level direction; all database traversal, hypothesis generation and candidate ranking were handled by the agents themselves.

During that search, one agent landed on something that stopped it in its tracks. While scanning raw DNA flanking an unusual reverse transcriptase gene, the agent wrote:

“[The DNA next to the RT] is spectacular: I can see by eye a tandem repeat array … that’s a CRISPR-like … repeat array?!”

The underlying reverse transcriptase had appeared in earlier studies of a jumbo phage, but no prior work had connected it to the adjacent repeat array or to a second, functionally unknown accessory protein sitting nearby. ART is defined by that trio: the enzyme, the partner gene, and an array holding between 3 and 21 copies of a short repeat sequence. Lab experiments confirmed the array is transcribed into a distinct set of short RNAs — a result consistent with, though not equivalent to, how CRISPR arrays produce guide RNAs. What those RNAs actually do remains an open question, and Anthropic is careful to say the biological function of ART has not been established.

One important caveat accompanies the announcement. When Anthropic ran the same discovery campaign 10 additional times to test reproducibility, none of the reruns caught the repeat array — every subsequent run missed the defining feature that made ART notable. The company attributes this to the sheer size of the search space and the inherently variable behavior of large multi-agent systems. The findings are published as a preprint and have not yet passed peer review.

A Different Kind of AI Biology Tool

Anthropic’s announcement lands in a life sciences AI landscape that is already crowded. Google DeepMind has the longest production track record: AlphaFold 2 reshaped protein structure prediction in 2021, AlphaFold 3 extended that work to protein-molecule and protein-nucleic-acid interactions in May 2024, and DeepMind has since released AlphaGenome Atlas, a publicly available database predicting the effects of roughly 9 billion possible nucleotide variations in the human genome. OpenAI, NVIDIA, Microsoft and Meta all market life sciences products as of 2026.

The critical distinction is in the mode of AI use. DeepMind’s tools are specialized, task-specific predictive systems trained to solve well-defined biological problems — more like a very accurate calculator for a specific domain. Anthropic’s approach deploys a general-purpose conversational agent to perform open-ended hypothesis generation and genome mining with minimal human guidance — closer to a junior scientist given a broad research question than a tool handed a structured input. Anthropic claims to be the first major AI lab to demonstrate this approach producing a potentially novel system validated with real wet-lab experiments. Its Bay Area laboratory operates at biosafety levels BSL-1 and BSL-2 and handles no human pathogens; all physical experiments are conducted by human scientists.

What This Means If You’re Studying Biology, CS or Data Science

The workflow Anthropic describes — having Claude survey literature, reproduce established results from public databases, generate candidate reports, and critically self-evaluate them — maps almost exactly onto what junior researchers, research assistants, and first- and second-year graduate students are typically asked to do. That overlap is worth sitting with.

Students who develop the skill to direct and evaluate AI agents performing genome mining will be in a strong position as this model of research spreads. Students whose primary contribution is manual database curation face compression of that role. Practically, the tools Anthropic used — Claude Science and Claude Code — are described in the announcement as available to any scientist today, meaning there is no barrier to experimenting with comparable workflows right now.

There is also a direct near-term opportunity. Anthropic’s life sciences group explicitly invites external collaboration, noting that if researchers have a proposal for an interesting biological question, the team wants to hear from it. For students or early-career researchers sitting on an interesting genomics dataset or a genomic anomaly they do not have the computational resources to pursue, that is a concrete door worth knocking on.

The ART discovery is preliminary, reproducibility is an acknowledged problem, and peer review has not yet happened. But the broader signal is hard to miss: autonomous AI agents are beginning to operate as participants in the earliest, most open-ended stage of biological research — not just as tools that answer questions humans already know how to ask.

Source: Anthropic