Claude agents spotted a CRISPR-like DNA pattern no one had characterized. Nobody knows yet what it does.

A familiar beginning: strange patterns and big discoveries

Some of the biggest tools in the history of biotechnology started the same way: someone noticed a strange pattern in DNA and asked why it was there. Restriction enzymes, which cut DNA at specific short sequences, were first spotted inside bacterial immune systems, where they shred the DNA of invading viruses. Scientists realized they could use those molecular scissors to cut and paste genes, and the modern biotechnology industry followed. Taq polymerase, the heat-tolerant enzyme behind PCR, came from a bacterium living in a Yellowstone hot spring; it now underpins much of medical diagnostics. And CRISPR, today the foundation of gene-editing medicines, began as nothing more than an odd, repeating sequence in bacterial genomes that researchers could not initially explain.

On September 23, 2026, Anthropic announced that this time, the noticing was done by an AI. The company says that Claude agents, searching DNA-sequence databases largely on their own, flagged a previously uncharacterized enzyme system in bacteriophages, the viruses that infect bacteria. Anthropic calls the system an array-associated reverse transcriptase, or ART. The name describes a structure: a reverse transcriptase gene, an enzyme that copies RNA into DNA, sitting next to a partner gene and a long, evenly spaced array of DNA repeats that resembles a CRISPR array.

That resemblance is the reason people are paying attention, and it is also where the honesty has to start. CRISPR arrays work as a kind of molecular memory bank, storing RNA sequences that let CRISPR-Cas systems be programmed to cut chosen DNA targets. Anthropic's system looks structurally similar, but nobody yet knows what ART actually does, or whether it can be programmed at all. This article separates what has been observed from what is still unknown and what is simply interpretation.

What Anthropic says happened

The story begins in spring 2026, when Anthropic formed a research group to test whether general AI models could systematize a process that has historically depended on individual scientific intuition: reading vast quantities of DNA sequence data and picking out the parts that look unusual. The company also opened a new life sciences laboratory in the Bay Area to test the hypotheses that the agents generate. According to the announcement, the lab operates only at biosafety levels 1 and 2, does not handle pathogens that infect humans, and all wet-lab work is performed by human scientists.

The typical workflow Anthropic describes goes like this: Claude is pointed at a protein family, reads the relevant literature, reproduces established results from public data to check its own methods, and then searches for family members or genomic neighbors that fit no described system. It writes a short, human-readable report for each candidate proposing a function and laying out the evidence. In follow-up analyses, Claude critically evaluates its own candidates, and according to Anthropic, most are eliminated at that stage. When a candidate survives review, human scientists express the protein in standard laboratory strains and characterize it biochemically and structurally, with Claude helping to interpret the data.

For the ART campaign, Claude agents were given a prompt to search a massive database of DNA sequences for interesting new examples of reverse transcriptases. Anthropic states that the agents gathered over 200,000 RTs, picked out 3,500 new candidate systems, and narrowed those to the 20 most compelling candidates, each summarized in a human-readable report. The search ran for about 21 hours, used roughly 950 agents and consumed about 210 million tokens. Anthropic says its own involvement was limited to the initial prompt and the laboratory work that followed.

The moment of detection

During the search, one agent flagged an unusual RT family and decided to look more closely at the raw DNA near the gene. Anthropic reports that the agent exclaiming over what it saw, quoting it as saying the DNA next to the RT was "spectacular" and that it could see "by eye a tandem repeat array," followed by "that's a CRISPR-like ... repeat array?!" Readers should treat that line the way it is presented: as reported by Anthropic, not as a quote that can be independently verified from outside the company.

The agent then proceeded much as a human scientist would. It counted the repeats and measured their spacing, compared the layout with known RT systems, searched the literature for any previous report of the pattern, and, after a thorough analysis, filed a report for human review.

The resulting system, ART, is found mainly in bacteriophages and consists of three parts: the reverse transcriptase, a partner gene beside it, and a long array of evenly spaced DNA repeats. Anthropic notes that the underlying RT, found in a jumbo phage, had been identified in previous studies, but that Claude appears to be the first to notice the system's defining features: the associated array of non-coding DNA sequences and the additional accessory protein of unknown function. In other words, the raw ingredient was already sitting in public databases; the contribution claimed here is the pattern recognition that connected the pieces.

Anthropic's first laboratory experiments show that the ART array is expressed as a set of distinct short RNAs, which is the same general behavior seen in CRISPR arrays, where the repeats are processed into small guide RNAs. That is suggestive, and Anthropic is careful to call it exactly that: the experiments suggest something analogous may be at play, and further experiments are underway to determine how ART works.

What is known, and what is not

Anthropic points out that the combination of features found in ART has only ever been seen together in a handful of other systems, all of which are programmable and perform operations like cutting, copying and pasting DNA. That is a meaningful observation, but it is an argument from analogy, not a demonstration. Knowing whether ART is programmable requires further work, and Anthropic says so explicitly. A structural resemblance to CRISPR does not mean ART is a CRISPR successor, and nothing in the announcement shows that it cuts, copies or pastes DNA in a controllable way.

There is also a structural fact about this story that deserves emphasis: every claim here comes from Anthropic itself. The discovery, the agent statistics, the laboratory confirmation that the array is expressed as short RNAs, and the framing of the significance are all reported by the company that also built the AI being celebrated. Anthropic released a preprint covering the work in more detail, which provides a technical account, but the piece could not be independently retrieved for verification at the time of writing. Independent replication by outside laboratories will be the real test, and none has been reported yet.

Anthropic did share one outside reaction. Feng Zhang, a pioneer of CRISPR genome editing and a professor at MIT and the Broad Institute, said after reviewing the preprint that it is an exciting example of how AI agents can contribute to biological discovery, that the identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation, and that he hopes the work encourages more scientists to explore how AI can support their research. That is a positive comment from a highly credible scientist, but it is a comment on an unreviewed preprint, not a validation of the system's function.

Why it matters, and what to watch

The deeper significance of the announcement may be less about ART itself and more about the process. Genome mining, the practice of searching sequence databases for uncharacterized genes and working out what they do, is how nearly all reverse transcriptase families have been found in recent years. For an expert scientist, Anthropic says, the kind of analysis its agents performed can take weeks to months of work. If agents can reliably triage hundreds of thousands of sequences down to a short list of candidates worth testing, that changes the economics of hypothesis generation, at least in principle.

Anthropic itself treats the hypotheses as an object of study. With hundreds to thousands of candidate reports from a single campaign, the team is asking what distinguishes the proposals worth testing from those set aside, and feeding what it learns back into the instructions it gives Claude. The company describes this as teaching the model to mimic human scientific taste.

There is a fair skeptical reading too. The expensive, uncertain part of biology is not usually spotting an odd pattern; it is working out what the pattern does, which still requires the human-run laboratory work Anthropic describes. The announcement is candid about this, presenting ART as an early result shared to demonstrate capability and give the community insight, not as a solved problem.

My view is that the honest summary sits in the middle. The observed facts are notable: an AI agent, working from a high-level prompt, surfaced a specific, testable structural pattern that previous studies of the same enzyme family had missed, and human scientists confirmed a CRISPR-like expression signature in the lab. The predictions, that ART may turn out to be a programmable DNA-editing system, remain unproven. And the interpretation, that this heralds a new mode of AI-accelerated discovery, is a plausible hypothesis that this single result cannot yet establish. If future work shows ART is programmable, this announcement will look like a genuine milestone in the lineage that gave us restriction enzymes, PCR and CRISPR. If it does not, the methodological lesson, that AI agents can flag candidate systems at scale for human testing, will still stand, just with quieter stakes. Either way, the next chapters belong to experiments, not to announcements.