Anthropic reports Claude AI identified novel CRISPR-like enzyme system in bacteriophages

Identification of the ART system

Anthropic announced on 23 September 2026 that its artificial intelligence model Claude identified an uncharacterized biological mechanism in bacteriophage DNA. The system, designated Array-associated Reverse Transcriptase (ART), was detected in viruses that infect bacterial cells. It consists of short, repeated DNA sequences arranged alongside two specific genes: one encoding a reverse transcriptase enzyme that transcribes RNA into DNA, and a second encoding a protein whose cellular function is currently unknown. Researchers noted that this repeating arrangement shares structural similarities with CRISPR, a natural bacterial defense mechanism adapted into therapeutic gene editing. While the reverse transcriptase gene had appeared in earlier sequencing catalogs, Anthropic reported that the AI model was the first to recognize that the adjacent components operate as an integrated system.

Autonomous screening architecture

To locate the pattern, Anthropic deployed 950 Claude software agents operating in parallel across genomic databases for 21 hours. The computational workload consumed approximately 210 million tokens and screened more than 200,000 reverse transcriptases, compressing an analytical process that typically requires weeks or months into less than a day. The software flagged 3,500 candidate sequences before selecting 20 top candidates for targeted experimental evaluation. Anthropic stated that human researchers set the initial prompts and parameters, while the AI agents autonomously chose analytical directions and evaluated scientific literature. Stanford University associate professor of bioengineering Stanley Qi commented on the computational capabilities demonstrated by the screening run.

Nature contains an enormous diversity of molecular systems we barely understand, and some patterns are very complicated but highly meaningful; AI could greatly expand our ability to explore them more effectively and rapidly, in this case, in just 21 hours.

Wet lab operations and biosafety

The discovery marks the first public result from Anthropic's molecular biology laboratory in the San Francisco Bay Area, which the company established in spring 2026. The facility allows researchers to test AI-generated biological hypotheses through physical experiments conducted by laboratory staff. Anthropic stated that research at the facility is restricted to low-level biosafety classifications and does not involve pathogens capable of infecting human hosts. Claude does not directly operate laboratory hardware, though company leadership noted automated execution under safety guardrails may be evaluated in the future. Anthropic chief executive Dario Amodei addressed both the therapeutic prospects and the safety considerations of frontier AI in biology during a videoconference briefing to the United Nations Security Council on 23 September 2026.

Its precise function, biotechnological utility (if any), or level of significance is not yet clear, but at minimum it is work I would have been proud to do as a PhD student.

Scientific validation and unresolved questions

Anthropic published the research as an unreviewed preprint, emphasizing that the exact biological function of the ART system remains unverified. Initial laboratory tests conducted by the company suggested the system could possess programmable characteristics, but researchers have not yet confirmed whether the enzyme is active in living cells. MIT professor and CRISPR pioneer Feng Zhang described the identification of the repeated DNA sequences as intriguing and noted that the findings warrant further investigation. Dario Amodei indicated that a separate research group at Stanford University independently documented a comparable enzyme arrangement. Other researchers, including Washington University School of Medicine microbiologist Kevin Blake, urged caution regarding comparisons to CRISPR until peer-reviewed experimental verification is completed.