Claude AI Agents Independently Discover a New System in Viral DNA That Resembles CRISPR

The Anthropic team has reported the first major success of its new biology lab. Autonomous agents powered by the Claude neural network were able to discover a previously undescribed molecular system called ART. It is found mainly in bacteriophages (viruses that infect bacteria) and has characteristics that make it potentially interesting for genetic engineering.

The main value of this discovery lies not only in the find itself but also in how it was made: the artificial intelligence analyzed a massive array of genomic data and formulated a hypothesis without step-by-step human oversight.

How 950 agents replaced months of work by bioinformaticians

The researchers gave Claude a high-level task: to find unusual systems linked to reverse transcriptases — enzymes that copy information from RNA into DNA — in open genome databases. The algorithms then built the entire search and filtering process on their own.

About 950 AI agents took part in the experiment. Over 21 hours of continuous work, they used 210 million tokens, assembling an initial database of 200,000 reverse transcriptases. The neural networks filtered this array on their own, identified 3,500 potentially new structures and narrowed the final list down to the 20 most promising candidates. According to Anthropic’s estimates, a human researcher would have spent anywhere from several weeks to several months on similar work.

During the work, one of the agents noticed an unusual transcriptase and the stretch of DNA adjacent to it. The neural network spotted the repeating array on its own, counted the number of repeats, measured the distances between them, checked the scientific literature on known biological systems and only then handed the candidate over to human biologists for lab tests.

What ART is and why the system is being compared to CRISPR

The structure the algorithm found consists of a reverse transcriptase, an unknown gene and a long array of evenly spaced repeating DNA sequences. It is the structure of the repeats that looks like the famous CRISPR arrays, which allow CRISPR-Cas mechanisms to find and precisely cut the desired sections of genetic material.

The first lab experiments, described in a preprint on AlphaXiv, confirmed that the ART array is indeed expressed as a set of individual short RNAs. This gives researchers grounds to suggest a similar programmable mode of operation. The few biological systems known to combine these properties are capable of complex DNA operations, including copying, cutting or moving its fragments.

However, Anthropic explicitly stresses that ART’s exact function remains a mystery. The neural network found the structure but has not yet proven its ability to perform targeted genome editing. Still, one of the pioneers of CRISPR technology, Feng Zhang of MIT and the Broad Institute, has already reviewed the results and called the combination of repeating RNA arrays with reverse transcriptases an extremely intriguing avenue for further study.

It is still too early to call ART a “new CRISPR.” For now, it is above all a historic precedent: AI models are moving beyond simply searching databases and becoming independent virtual lab assistants capable of spotting structural anomalies. Whether the discovered system proves useful in real biotechnology or medicine will be shown by further in vitro tests.