In late September 2026, the artificial intelligence company Anthropic said that its model Claude had autonomously discovered a previously uncharacterized enzyme system with CRISPR-like repeats in bacteriophage DNA. The company named the system array-associated reverse transcriptases, or ART, and described it as the first result from its new life sciences laboratory in the Bay Area. The finding, announced on 23 September 2026, arrived after a 21.5-hour search that used roughly 950 Claude agents working through 215.6 million tokens. Anthropic released a technical report and a preprint that has not been peer reviewed.
Anthropic is known for building the Claude family of AI models, and it has recently expanded into biological research. The new lab does only BSL-1 and BSL-2 work, does not handle pathogens that can infect humans, and relies on human scientists for all laboratory experiments. The company framed the ART discovery as a demonstration of what autonomous AI agents can do when pointed at a massive DNA database. The prompt given to Claude was a high-level brief to search for new reverse transcriptase systems, and the company says its scientists were involved only in that initial prompt and in subsequent lab work.
Reverse transcriptases are enzymes that copy RNA into DNA, a trick used by several viruses and bacteria. CRISPR, by contrast, is a gene-editing tool derived from an ancient bacterial defense mechanism. CRISPR arrays are repeating DNA sequences that, together with cas genes, help bacteria remember and cut invading genetic material. The ART system shares some structural resemblance to CRISPR arrays, which is why the comparison has drawn attention. But ART has no cas genes nearby, and its actual function remains unknown.
The announcement comes as AI companies increasingly turn to scientific research to prove the value of their most capable models. Anthropic is preparing to go public, and the company has said it wants to attract more scientists to its lab and expand into areas such as drug discovery. The ART result is an early test of whether AI agents can make genuine contributions to biology, or whether they are better understood as tools that accelerate human researchers. The preprint and the company's blog post have not been peer reviewed, and independent experts have cautioned that the discovery is preliminary.
Key Facts
Anthropic announced on 23 September 2026 that about 950 Claude agents, running for 21.5 hours across 215.6 million tokens, autonomously discovered the ART system. The agents searched a database of 1.9 billion protein clusters, recovered roughly 200,000 reverse transcriptases, and scored 3,564 candidate partner families. They filed 19 reports for human review. The campaign used Claude Mythos 5, with one agent planning and running each task while a second reviewed it. The agents narrowed the candidates from about 200,000 down to 20 of the most compelling systems. Anthropic says this analysis can take weeks to months of work for an expert scientist.
The Next Web reported on 23 September 2026 that ART has three parts: a reverse transcriptase, a partner gene beside it, and an array of evenly spaced DNA repeats. ART arrays hold 3 to 21 copies of a short repeat, and none has the cas genes that CRISPR systems carry nearby. In lab tests, Anthropic's scientists found the array is read out as a set of distinct short RNAs. In published data from a Staphylococcus phage, those RNAs made up as much as 8 percent of the phage's RNA 15 minutes after infection. The preprint says the team has not yet shown the enzyme is active or that it works on these RNAs.
The Verge reported on 23 September 2026 that Claude found the enzyme system after searching a massive database of DNA sequences, and that one agent spotted an unusual repeating pattern and flagged it for human review. The Verge's Robert Hart noted that it remains unclear whether the discovery will have any practical applications, let alone be as transformative as CRISPR. The company said it felt it was important to share such findings early, both to demonstrate Claude's capabilities and to give the broader community insight into its work.
Feng Zhang, a CRISPR pioneer and professor at MIT and the Broad Institute, reviewed the preprint and said the identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation. Decrypt reported on 23 September 2026 that even Anthropic CEO Dario Amodei admits the system's precise function, biotechnological utility if any, or level of significance is not yet clear. Amodei also said that at minimum it is work he would have been proud to do as a PhD student. The find was hard to repeat: Anthropic ran the same campaign ten more times and all ten missed the array. In fixed tests, its four most capable models described the array in at least 90 percent of attempts when given the DNA directly, but with files and tools the rate fell as low as 32 percent.
Anthropic says the underlying reverse transcriptase had been identified before, but Claude appears to be the first to notice the system's defining features: an associated array of non-coding DNA and an accessory protein of unknown function. The system is found mainly in bacteriophages, the viruses that infect bacteria. The lab work was performed by human scientists, and the company says work to understand the primary function of ART is ongoing.
Analysis
The bigger picture here is that Anthropic has demonstrated a new kind of scientific workflow, one in which a large fleet of AI agents can comb through billions of protein clusters and surface a candidate that human experts then review. The company is careful to say that its scientists were involved only in the initial prompt and in lab work, but the scale of the search is the point. No human team could read 1.9 billion protein clusters in 21.5 hours, and the agents recovered about 200,000 reverse transcriptases along the way. The finding is not a cure or a product. It is a signal that autonomous agents can generate hypotheses at a pace that changes what a research group can attempt.
Yet the judgment must be tempered by the fact that ART's function is unknown. Anthropic itself says it does not know what the system does, whether it is useful, or whether it is important. The comparison to CRISPR is tempting because the arrays look similar, but CRISPR's power came from a decades-long effort to understand and then engineer a bacterial defense system into a gene-editing tool. ART has no cas genes, and the preprint has not shown that the enzyme is active or that it acts on the short RNAs. The discovery is a starting point, not a conclusion.
What this really means is that the value of the ART result may lie less in the enzyme itself than in the method that found it. Anthropic ran the same campaign ten more times and all ten missed the array, which suggests the discovery depended on a specific path through the search space. That is a warning about reliability. In fixed tests, the company's four most capable models described the array in at least 90 percent of attempts when given the DNA directly, but with files and tools the rate fell as low as 32 percent. The gap between a clean test and a real research environment is where the hard engineering work remains.
The company's decision to release a preprint and a technical report before peer review also invites scrutiny. Academic norms would typically require more validation before a claim of autonomous discovery. Anthropic says it wanted to share early findings to demonstrate Claude's capabilities and to give the community insight. That is a reasonable goal, but it also serves a commercial purpose as the company prepares to go public. The scientific community will need to reproduce the result and determine whether ART is a genuine biological novelty or an interesting pattern that requires much more work to understand.
Why It Matters
If AI agents can reliably surface new biological systems, the implications for drug discovery and biotechnology are significant. Reverse transcriptases are already important in research and medicine, and a new family of RNA-linked systems could offer fresh tools for editing or sensing genetic material. But the more immediate impact may be on how science is funded and organized. Anthropic's lab is betting that AI can shorten the time from database to hypothesis, and the ART campaign is a proof of concept. The company says the analysis that led to the 20 most compelling candidates can take weeks to months for an expert scientist. Compressing that into 21.5 hours changes the economics of early-stage research.
The result also matters because it tests the boundary between human and machine discovery. Claude agents did the searching and the flagging, but human scientists did the lab work and will do the follow-up. The system was found mainly in bacteriophages, viruses that infect bacteria, and its arrays hold 3 to 21 copies of a short repeat. Those are concrete details that other labs can check. The preprint's admission that the enzyme's activity has not been shown is a sign of appropriate caution. The scientific value will be determined by whether independent groups can confirm the array and figure out what it does.
There is also a competitive dimension. AI companies are increasingly using scientific results to prove the value of their models, and Anthropic's announcement comes as it seeks to attract more scientists and expand into areas like drug discovery. The Verge noted the contrast with OpenAI's approach to mathematics. For Anthropic, the ART finding is both a research milestone and a marketing moment. For biologists, it is a new puzzle. For the public, it is a reminder that AI's most consequential contributions may arrive quietly, as a pattern in a database that only a machine had the patience to notice.
Next Up
Anthropic says work to understand the primary function of ART is ongoing, and its role is not yet known. The company has released a technical report and a preprint, and independent researchers are expected to examine the data, attempt to reproduce the array, and test whether the enzyme is active. Feng Zhang's comment that the finding is genuinely intriguing and merits further investigation suggests that at least some experts see a reason to look closer. The next step is laboratory validation, not another AI campaign.
Anthropic also faces questions about the reliability of its agentic search. The fact that ten reruns missed the array, and that tool use dropped the description rate to as low as 32 percent, means the company will need to improve consistency before making broader claims about autonomous discovery. The lab, which does only BSL-1 and BSL-2 work and does not handle human pathogens, will continue to rely on human scientists for experiments. Whether ART becomes a useful biotechnology or remains a curious pattern in phage DNA will depend on the slow, careful work that follows the headline.
Comments (0)
Log in or sign up to leave a comment.
No comments yet. Be the first to share your thoughts.