Anthropic's new life sciences research group says an autonomous run of 950 Claude agents found a previously undescribed enzyme system hiding in bacteriophage DNA -- and, seven days after the group's own Life Sciences Verification Program opened for beta applicants, this is its first disclosed result. The system, which Anthropic named array-associated reverse transcriptases, or ART, consists of a reverse transcriptase enzyme, a neighboring gene of unknown function, and a long array of evenly spaced DNA repeats -- a layout that resembles the guide-RNA array at the heart of CRISPR.
The reverse transcriptase itself wasn't new -- it had already been identified in so-called jumbo phages. What Claude's agents found was the defining pattern around it: the repeat array sitting next to it, and the family resemblance to a system built to be programmable. Early experiments cited by Anthropic show the array expresses as distinct short RNAs, the same basic design CRISPR-Cas systems use to aim at a target sequence.
Why the resemblance matters: CRISPR's value was never the bacterial immune system it evolved from -- it was that scientists could reprogram its guide array to target almost any DNA sequence they chose, turning a bacterial defense mechanism into an editing tool used across medicine and agriculture. A second system with the same programmable-array structure, built around a different enzyme, would be a genuinely new lever on the same kind of problem -- if ART turns out to work the way its layout suggests. That "if" is doing all the load-bearing work in this story.
The autonomous run, in numbers
- 950 · agents
- Deployed in parallel across the run
Includes: Every agent instance working the 21-hour window
Excludes: Not sequential -- concurrent, not 950 separate runs stacked end to end - 210M · tokens
- Consumed over the full run
- 200,000+ · reverse transcriptases
- Candidate enzymes screened from the public sequence database
- 20 · finalists
- Narrowed down from 3,500 candidate systems
Of the 3,500 candidate systems that search turned up, Anthropic's agents narrowed the field to 20 finalists worth a closer look -- and ART is the one the company chose to describe in detail. 210M (tokens spent before anyone knew what, if anything, the finding actually does) That's the point Anthropic itself won't skip past. Chief executive Dario Amodei has said ART's "precise function, biotechnological utility (if any), or level of significance is not yet clear." The lab handles only BSL-1 and BSL-2 materials and no human pathogens, and the finding has been posted as a preprint -- meaning it has not yet passed independent peer review. Feng Zhang, the MIT and Broad Institute researcher who helped pioneer CRISPR gene editing, reviewed the finding and called it "an exciting example of how AI agents can contribute to biological discovery" -- notably a comment about the process, not a confirmation of what ART does.
What's nailed down versus what's still open
- ART is a genuinely novel enzyme system, distinct from previously catalogued CRISPR-family machinery.
- ART functions as a programmable, CRISPR-like targeting system.
- This result shows AI agents can now do independent scientific discovery, not just accelerate existing workflows.
That third line in the scorecard is the one worth reading most skeptically, because it's the one doing the most public-relations work.
The strongest case against the 'AI does science now' read
The market noticed regardless of the hedging. Some publicly traded gene-editing companies' shares fell on the news, a reaction that reads the finding as competitive -- a faster, cheaper way to find the next CRISPR-adjacent tool -- rather than as the early-stage, function-unknown research result Anthropic itself describes. That gap, between how the finding was priced and how carefully Anthropic itself described it, is worth sitting with before assuming this changes anything about who owns the next gene-editing platform. A stock move is a bet on what other traders will believe by tomorrow; it is not a peer review.
Anthropic formed the life sciences research group in spring 2026, built around a specific bet: that an AI system searching a public sequence database exhaustively, at a scale and patience no graduate student has, would surface patterns humans had walked past. This is the group's first public test of that bet, run inside a lab that deliberately keeps to BSL-1 and BSL-2 materials -- no human pathogens -- so the discovery pipeline itself could be demonstrated before it's ever pointed at anything higher-stakes.
The timing lines up with a second Anthropic life-sciences move the same week: the company opened its Life Sciences Verification Program to beta applicants on Sept. 17, a credential-based system that lets vetted biology researchers and drug-discovery teams turn off Claude's safety refusals for legitimate work, replacing real-time blocking with after-the-fact monitoring of their own retained traffic. That program arrived eight days after Anthropic disclosed five separate cases of users trying to get Claude's models to help with viral modification and toxin redesign using VPNs and burner accounts -- the same week's other reminder that a model capable of genuine biology research is also a model biosecurity researchers are actively watching for misuse.
Read together, the two moves describe one bet made by one company in one week: that Claude is now a genuine research instrument in biology, not just a chatbot that happens to know some biochemistry -- with this story's caveats about function sitting right next to that story's caveats about misuse.
- 950 Claude agents screened 200,000-plus reverse transcriptases over 21 hours, flagging a new system: ART.
- ART's DNA-repeat layout resembles a CRISPR array, and early data shows it expresses as short RNAs.
- CRISPR pioneer Feng Zhang called the finding worth investigating; it's a preprint, not peer-reviewed.
- Some gene-editing company stocks fell on the news, reading it as competitive rather than collaborative.
- Caveat: Anthropic's own materials say ART's actual function -- if it has one -- is not yet known.