Anthropic Says Its Claude AI Agents Autonomously Found a New CRISPR-Like Enzyme System

Anthropic said this week that autonomous AI agents built on its Claude models spent 21 and a half hours searching a database of nearly 1.9 billion protein clusters and surfaced a previously uncharacterized enzyme system in the DNA of bacteriophages, the viruses that infect bacteria. The company is calling the system array-associated reverse transcriptases, or ART, and says its layout of evenly spaced DNA repeats resembles the architecture that makes CRISPR gene-editing tools programmable.

The finding, detailed in a post published Wednesday on Anthropic's website and in an accompanying preprint by Yoon and colleagues, marks the first public output of a life sciences research group Anthropic formed this spring. The company is careful to note what it has not shown: nobody yet knows what ART actually does, whether the enzyme is active, or whether it has any practical use.

A 21-Hour Search With No Human in the Loop

According to the preprint, Anthropic's scientists gave the Claude-based agents a single research brief: look for new reverse transcriptase systems, enzymes that copy RNA into DNA, across a database of roughly 1.9 billion protein clusters. From there, the agents worked without direct human intervention. One agent would plan and carry out a task, a second would review the work, and the system opened new lines of inquiry as findings warranted, according to Anthropic's account.

Over 949 agent sessions and 215.6 million tokens, the agents assembled roughly 200,000 reverse transcriptase sequences, narrowed those to 3,564 candidate partner-gene families, and wrote up 19 reports on the most compelling leads. ART emerged as the standout: a reverse transcriptase paired with a neighboring gene of unknown function and a long, evenly spaced array of DNA repeats, a structural echo of the CRISPR arrays that store the RNA guides used in gene editing. Outside coverage of the announcement, including reporting from Gizmodo, notes the system turns up largely in so-called jumbo bacteriophages that infect Staphylococcus bacteria.

Anthropic said the underlying reverse transcriptase itself had shown up in earlier genomic surveys, but that no one had previously flagged its pairing with the repeat array and the accessory gene, the combination that gives ART its CRISPR-like shape. The preprint is explicit that this is a hypothesis pending confirmation, not a validated discovery: the authors write that they have not yet demonstrated the enzyme is active or that it interacts with the RNAs it is theorized to use, and the paper has not been through peer review.

Where the Work Happens, and Who Is Doing It

Anthropic's life sciences group operates out of a Bay Area laboratory that it says runs at biosafety levels 1 and 2, the two lowest tiers in the federal containment framework maintained by the CDC and NIH, and does not handle pathogens capable of infecting humans. Every physical experiment described in connection with the ART finding, the company said, was carried out by human scientists rather than by Claude itself. Chief executive Dario Amodei said Anthropic is not currently allowing its models to run laboratory experiments autonomously, though he suggested that could eventually change under supervision. The group's earlier work, according to Anthropic, focused on more conventional computational biology, including tracing the evolutionary history of CRISPR systems and searching for enzymes with potential use in cell and gene therapies, groundwork the company says fed directly into the pipeline that produced ART.

The disclosure lands amid a broader push by Amodei to position AI as a near-term force in biology and medicine; he has said publicly that he expects AI to help cure most diseases within five to ten years. He has also named bioterrorism as one of his central worries about advanced AI, a tension that runs through Anthropic's own framing of the ART announcement: the same automated search that can flag a potentially useful biological pattern could, in principle, be pointed at something harmful. Coverage from Al Jazeera situates the announcement inside that wider, ongoing debate over how AI companies police dual-use biological research.

Scientists Split Between Intrigue and Caution

Reaction from outside researchers has been mixed. Feng Zhang, the MIT and Broad Institute scientist known for his early work on CRISPR, called the finding "genuinely intriguing" and said it merits further investigation, framing it as a demonstration of how AI agents can contribute to biological discovery. Stanford bioengineer Stanley Qi described it as "incredibly exciting," pointing to the system's ability to recognize an unusual biological pattern in a matter of hours rather than the months or years such surveys have traditionally taken.

"There's nothing to indicate this is a rival to CRISPR-the-technology, or could be developed into any practical application," said Kevin Blake, a microbiologist at Washington University.

Dimitri Perrin, a researcher at Queensland University of Technology, drew a similar distinction, saying ART is CRISPR-like in its architecture but that there is no evidence it is CRISPR-like in its function. Even Amodei, in the company's own materials, acknowledged that ART's "precise function, biotechnological utility, if any, or level of significance is not yet clear." Coverage from the Digital Watch Observatory noted that researchers outside Anthropic have raised the possibility of apophenia, the risk that a pattern-seeking system surfaces a structure that does not hold the significance later ascribed to it.

What Happens Next

Anthropic said further wet-lab experiments aimed at determining what ART actually does are already underway, and it is inviting outside scientists to propose their own studies of the system. The company published the preprint openly rather than waiting for journal review, a choice it framed as an effort to let independent researchers scrutinize the agents' methodology and, potentially, replicate it against their own genomic databases.

For the biotechnology field, the more consequential question may not be about ART itself but about the search process that produced it. If an agentic pipeline built on a general-purpose chatbot can comb billions of protein sequences and flag a structurally novel candidate in under a day, without a human choosing which sequences to examine, that capability is likely to be tested again, by Anthropic's own group and by rivals and academic labs experimenting with similar agentic approaches to genomic research. Whether ART turns out to be a scientific footnote or the first entry in a longer list will depend on laboratory results that, by Anthropic's own account, do not yet exist.