Anthropic's Wet Lab Unveils ART: Claude Discovers a CRISPR-Like Enzyme System

In what marks the first major physical science discovery spearheaded by an autonomous LLM swarm, Anthropic has revealed that Claude discovered a novel, uncharacterized biological system in bacteriophage DNA with molecular features reminiscent of CRISPR.

The discovery—named Array-associated Reverse Transcriptases (ART)—was identified after an autonomous swarm of roughly 950 Claude agents mined genomic datasets for 21 continuous hours, consuming 210 million tokens. Following computational identification, human biochemists inside Anthropic’s newly established Bay Area wet lab experimentally verified that the ART repeat array is transcribed into distinct, short non-coding RNAs in living cells—a foundational hallmark of programmable molecular machines.

Here is how Anthropic built an end-to-end biological discovery pipeline, what Claude actually found in jumbo phage genomes, and why this represents a seismic shift for AI-driven therapeutics.

The Architecture: Swarm Genome Mining at Scale

Traditional bioinformatic pipelines rely on heuristic sequence alignments (like BLAST or HMMER) designed around known protein structural motifs. While effective, these pipelines often miss novel functional systems where components do not match predefined homology models.

Anthropic structured the search as an autonomous multi-agent exploration workflow with minimal human intervention:

Broad Candidate Mining: Claude agents combed through massive public and proprietary metagenomic databases, collecting and categorizing more than 200,000 reverse transcriptases (RTs).
Hypothesis Generation & Filtering: The agents autonomously triaged the dataset down to 3,500 candidate systems exhibiting unusual genomic neighborhoods, ultimately drafting deep-dive dossiers on 20 high-priority targets.
De Novo Contextual Analysis: While examining the genetic flank of an uncharacterized RT within a giant "jumbo phage" virus, one Claude agent noticed an anomaly: an upstream non-coding region containing regularly spaced, 200-nucleotide tandem DNA repeats paired with an uncharacterized accessory protein.
Literature & Phylogeny Cross-Referencing: The agent autonomously counted the repeat units, computed inter-repeat spacing, conducted phylogenetic mapping against known RT families, and queried PubMed to verify that this structural architecture had never been documented in biological literature.

What would typically take an interdisciplinary team of bioinformaticians months of iterative manual screening was completed in 21 hours of compute time.

What Claude Found: The ART Molecular Machine

To understand why molecular biologists are paying close attention, it helps to understand how CRISPR works. In nature, CRISPR-Cas systems act as adaptive bacterial immune systems: bacteria store fragments of viral DNA inside repeating arrays, transcribe them into short guide RNAs (crRNAs), and use Cas nucleases to cut matching viral sequences.

Claude’s discovery, ART, combines three distinct biological components within bacteriophages:

A Phage Reverse Transcriptase (RT): An enzyme capable of converting RNA back into DNA, traditionally used by retroviruses and retrotransposons.
A Tandem Repeat Array: A flanking genomic array of regularly spaced DNA repeats that Anthropic’s wet lab proved is actively transcribed into discrete, short RNA fragments.
An Accessory Partner Gene: A conserved protein of unknown biochemical function located immediately adjacent to the RT cassette.

Only a tiny handful of known biological systems share this precise architectural triad—and virtually all of them (including CRISPR-Cas, Retrons, and Bridge Recombinases) operate as programmable molecular machines capable of targeted genetic cutting, copying, or sequence-specific insertion.

"This is an exciting example of how AI agents can contribute to biological discovery. The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation."
— Dr. Feng Zhang, CRISPR Pioneer and Professor at MIT and the Broad Institute

Moving from In Silico Prediction to the Wet Lab

Unlike pure computational biology startups that rely entirely on predictive scoring benchmarks, Anthropic established a physical molecular biology lab in the San Francisco Bay Area in spring 2026. Operating under Biosafety Level 1 and 2 (BSL-1/BSL-2) containment protocols, human molecular biologists test and validate the hypotheses generated by Claude.

When Claude flagged the candidate ART locus, human scientists synthesized the sequence and assayed its transcription in bacterial hosts. The experiments confirmed that:

The repeat array is actively transcribed and processed into individual short RNA fragments, confirming the agent's prediction that the non-coding array was not evolutionary junk DNA.
The phage-derived RT expresses stably alongside its accessory protein, setting the stage for functional assays to determine whether ART can execute targeted RNA-templated DNA synthesis or novel gene-editing operations.

Anthropic emphasized that while Claude designs experiments, models structures, and analyzes sequencing runs, all physical pipetting and hazardous material handling remains strictly under human operator control.

The Dual-Use Frontier: Cures vs. Biosecurity

This discovery arrives at a pivotal moment in the AI industry. Just weeks prior, Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman called on regulators and peers to adopt formal safety frameworks to prevent rogue autonomous agents from interacting with hazardous biological tools. Amodei has frequently stated his belief that AI agents will "cure most major human diseases within the next 5 to 10 years," while simultaneously warning of severe biosecurity risks if frontier models gain unconstrained wet lab autonomy.

By open-sourcing the pre-print and publishing the ART genetic architecture, Anthropic is positioning Claude not merely as a chatbot or coding assistant, but as a primary scientific discovery engine for fundamental genomics.

What to Watch Next:

Biochemical Characterization: Peer-reviewed wet lab assays will determine whether ART acts as a phage defense system against host bacteria or an active gene insertion mechanism.
Therapeutic Programmability: If the short repeat-derived RNAs can be engineered to guide the reverse transcriptase to custom DNA targets, ART could form the foundation of a new class of prime-editing or RNA-directed DNA writing tools.
Autonomous Closed-Loop Labs: Anthropic has hinted at eventual closed-loop integration with automated liquid-handling robotics, creating continuous hypothesis-experiment-refinement cycles for drug discovery.