Anthropic says its Claude model autonomously discovered a previously unknown enzyme system in bacterial viruses, one with structural features seen only in a handful of known systems, all of which turned out to be programmable gene-editing tools.
The company announced the finding on September 23, the first result from the molecular biology lab it opened in San Francisco this year. The system, named array-associated reverse transcriptases, or ART, sits in the DNA of bacteriophages, the viruses that infect bacteria. Anthropic is careful about what it claims: the function of ART is unknown. What caught attention is its shape. The enzyme pairs with a long array of evenly spaced DNA repeats, a pattern reminiscent of CRISPR, the bacterial immune system that scientists converted into the gene-editing platform now used across medicine and research.
How the discovery happened
Anthropic scientists gave the system a broad prompt: search public DNA databases for uncharacterized protein families associated with reverse transcriptases, enzymes that read RNA and write DNA, the reverse of the usual flow of genetic information. From there the work ran largely without human steering.
Roughly 950 Claude agents combed through the data for 21 hours, consuming about 210 million tokens. They identified more than 200,000 genes for reverse transcriptase enzymes and narrowed the field to 20 strong leads. One agent flagged something odd: a repeating pattern of DNA sequences sitting next to the gene for an unusual reverse transcriptase. Human scientists then took over, running bench experiments at the lab to confirm the system was real and active.
The division of labor matters. Anthropic stresses that all lab work was performed by human scientists, and the research involved only biosafety level 1 and 2 protocols, with no handling of pathogens that infect humans. The agents did the reading and hypothesis generation, work that would traditionally take a postdoctoral researcher months of database trawling.
Why the CRISPR comparison
Repeat arrays next to enzyme genes are how biology stores memory of past infections. In CRISPR, bacteria keep snippets of viral DNA in such arrays and use them to recognize and cut matching invaders. A repeat array next to a reverse transcriptase suggests a similar logic operating through RNA intermediates, which is why Anthropic describes the system as having properties reminiscent of CRISPR rather than being CRISPR.
The company notes that the combination of features found in ART, an enzyme plus a partner gene plus an ordered repeat array, has only ever appeared together in a small set of known systems. Every one of those systems is programmable and performs operations on DNA: cutting, copying and pasting genetic sequences. That history is the reason the finding generated attention beyond a routine database paper. A new programmable system would not automatically become a gene-editing tool, but the precedent is that such systems can be turned into one.
Many discoveries that have revolutionized biology and medicine started with a scientist noticing something odd in the staggering diversity of molecular machines found in nature.
Anthropic included that line in its announcement, positioning the work in the tradition of basic biology rather than product development.
Skeptics and caveats
Independent scientists quoted in coverage from The Straits Times and other outlets called for further verification before treating the claim as settled. The finding has not yet appeared in a peer-reviewed journal, and Anthropic has a commercial interest in showing its models doing work no other system can. The company published the sequence details and its methods, which gives outside labs what they need to replicate the search and the bench work.
The timing also drew notice. Anthropic chief executive Dario Amodei recently called for slowing AI development over safety concerns, and the announcement landed amid a global debate about safeguards for frontier models. The company framed the lab as an example of AI accelerating basic science under human control, with low biosafety risk and human hands on every experiment.
The wider context is a race among AI labs to demonstrate scientific discovery, not just benchmark scores. Google has pointed to materials science results from its models, and OpenAI has flagged gains in cybersecurity research. A biology discovery carries particular weight because bench validation, not just a leaderboard score, separates a real result from a plausible one. It is also the kind of claim that ages quickly: either independent labs confirm the activity of ART within months, or the announcement joins a long list of AI science demos that never replicated.
The economics behind the lab
Anthropic opened the San Francisco lab as part of a bet that frontier AI companies will need in-house wet labs to convert model output into validated discoveries. The economics reflect that bet. Database search is cheap in compute terms but generates hypotheses at a volume no human team can triage. Bench validation is the expensive bottleneck, measured in scientist time and reagent costs. The company structure, agents for search, humans for experiments, allocates each resource to the stage where it is least replaceable.
Competitors have taken different routes. Some license academic labs, others partner with pharmaceutical companies that already run discovery pipelines. Building an internal lab is the most expensive option and the one that gives a company the most control over what gets tested and how fast. If ART validates, expect the internal-lab model to spread quickly across the industry.
What comes next
Anthropic says the next phase is characterizing what ART actually does. If the system proves to be a programmable immune mechanism in phages, the practical applications could take years to develop, and gene-editing tools built on it would face the same delivery and safety questions that CRISPR therapies navigate today. If it turns out to be something narrower, the result still stands as a demonstration of scale: 950 agents, 21 hours, one candidate worth testing, confirmed at the bench.
Sources: Al Jazeera; The Straits Times; Anthropic (company announcement, September 23, 2026); India Today