- Anthropic had Claude run roughly 950 parallel AI agents for 21 hours (210 million tokens) to scan over 200,000 reverse-transcriptase sequences, narrowing 3,500 candidates down to 20 for deep review.
- The search turned up a previously unknown enzyme system — Array-Associated Reverse Transcriptases (ART) — in bacteriophage DNA, with a CRISPR-like repeat array that may be programmable.
- CRISPR pioneer Feng Zhang (MIT/Broad Institute) called it “an exciting example of how AI agents can contribute to biological discovery.”
- CEO Dario Amodei says humans still run every physical experiment — Claude does not yet control lab equipment, though he says that “eventually it may” be possible.
On September 23, 2026, Anthropic said its in-house biology lab had made a real discovery — and that Claude, not a human researcher, did most of the work. Over 21 hours, roughly 950 Claude agents combed public sequence databases and surfaced a new enzyme system hiding in the DNA of bacteriophages, the viruses that infect bacteria. It’s a small, technical finding on its face. But it lands one week after Anthropic’s own CEO was publicly warning about AI-enabled bioweapons risk — which is exactly why it’s worth a closer look.
What Claude Actually Found Hidden in Virus DNA
A CRISPR Look-Alike With an Unknown Job
The system Claude identified has been named Array-Associated Reverse Transcriptases, or ART. It has three parts: a reverse-transcriptase enzyme, a partner gene sitting next to it, and a long, evenly spaced array of repeating DNA. That last piece is the interesting one — the repeat structure looks structurally similar to a CRISPR array, the programmable “storage bank” of RNA sequences that makes CRISPR-Cas gene-editing tools work. Early lab testing showed the ART array is also expressed as a set of distinct short RNAs, which hints at a similar programmable function. Nobody, including Anthropic, knows yet exactly what ART does in the phage’s life cycle. Dario Amodei has noted that Stanford researchers previously described a system “in some ways similar” to this one, so Claude’s find builds on real prior science rather than appearing from nowhere.
Trend Insight — Every major CRISPR-adjacent discovery so far has expanded the gene-editing toolkit. If ART turns out to be programmable, it could become the next biotech platform — and the fact that an AI found the candidate first, not a grad student’s hunch, is the actual story.
How a 950-Agent AI Swarm Beat Years of Manual Search
From 200,000 Candidates to 20, in Under a Day
Anthropic gave Claude minimal starting instructions and let it work the problem end to end. The agents pulled more than 200,000 reverse-transcriptase sequences from public databases, flagged roughly 3,500 as structurally unusual candidates, and then independently ran literature reviews, tried to reproduce known results, and wrote up human-readable reports arguing for or against each candidate’s novelty. That process cut the list to 20 genuinely compelling systems for Anthropic’s human scientists to examine in the wet lab — the step that ultimately confirmed ART. The whole computational phase took 21 hours and about 210 million tokens, work that would traditionally take a research team a lot longer to grind through manually.
Trend Insight — This is the pattern to watch across every science-focused AI lab right now: agents don’t replace the scientist, they replace the triage. The bottleneck in biology research has always been narrowing millions of candidates to a shortlist worth testing — and that’s precisely the step AI agents are getting good at.
The Awkward Timing: Safety Warnings Meet a Robot-Assisted Biology Lab
Humans Are Still Holding the Pipette — For Now
Anthropic has spent much of 2026 warning publicly that frontier AI models could eventually help bad actors design bioweapons, and it built much of its safety messaging around that exact risk. So a research group running Claude agents through a real, physical biology lab — established in the Bay Area this spring, operating at biosafety levels 1-2 with no human pathogens — is a genuinely awkward juxtaposition, and outlets covering the story flagged it as such. Anthropic’s own framing is that the boundary still holds: every physical experiment is run by a human scientist, not the model. Amodei has said that “eventually it may even be possible for Claude itself to safely perform the experiments by autonomously controlling lab equipment… but we aren’t doing that today.” For now, Claude’s role stops at the screen.
Trend Insight — Watch this line, not the enzyme. “Claude designs the search, a human runs the experiment” is the exact boundary every AI safety framework leans on right now. The moment any lab moves that boundary — even one step, even for a routine task — is the moment this story changes from a science win to a safety story.
Related
- 60 AI Agents Just Attacked Math’s 150-Year Mystery
- Anthropic Hid Something in Every Word Claude Writes
- Anthropic Just Revealed Who Actually Builds Claude Now
- Nine Mathematicians Just Got a Seat. Not a Vote.
Sources
- Anthropic — “Claude discovers a novel enzyme system with CRISPR-like repeats” (Sep 23, 2026)
- TechCrunch — “Anthropic says its biology lab has already found something big” (Sep 23, 2026)
AI Biz Insider · AI Trends EN · aibizinsider.com
답글 남기기