Anthropic says its wet lab has already produced a significant discovery

Anthropic confirmed this week that the wet lab it set up in the San Francisco Bay Area last spring has identified an entirely new enzyme system the company describes as having "CRISPR-like properties." CRISPR, for anyone not tracking the field, is a natural bacterial immune system that became the standard gene-editing tool in labs worldwide. Anthropic's system, according to the announcement, was discovered inside the DNA of bacteriophages — viruses that infect bacteria — and can perform the same cut, copy and paste operations on DNA as CRISPR.

How novel it actually is remains an open question. Chief executive Dario Amodei acknowledged on X that the finding builds on prior work, and that a Stanford team had already identified a system "in some ways similar to what Claude found." The broader scientific community will need to validate whether the discovery is genuinely breakthrough or a variation on a known theme. Anthropic has not yet published a peer-reviewed paper; the information comes solely from company statements, a detail worth keeping in mind when reading the headlines.

The technically interesting part is the speed. According to Amodei, the discovery was made "mostly, though not entirely, by Claude," Anthropic's language model. The system ran roughly 950 agents in parallel, consumed about 210 million tokens, and completed the search in 21 hours of concentrated compute. A lab only months old producing a result like that in that timeframe is a data point that is hard to ignore, even if it should be taken with a grain of salt until formal publication.

The announcement arrives weeks after Amodei and other industry CEOs publicly called for slowing development and formalising safety-testing procedures, arguing models have become sufficiently dangerous. Amodei himself has said one of his biggest fears is AI-enabled bioterror, while in the same breath estimating AI will "cure most diseases within 5-10 years." Anthropic has clearly decided the benefit outweighs the risk and built an active lab.

Humans still do the wet work. Despite the capabilities Claude demonstrated in search and planning, the physical experiments were carried out by human scientists. The facility operates at biosafety levels BSL-1 and BSL-2 only and does not handle pathogens that infect humans. "We don't let Claude control lab equipment today," Amodei clarified, though he did not rule out full automation in the future with appropriate safeguards.

Anthropic is far from the only player here. Stanford researchers recently published a paper on using large language models with CRISPR; UCSF researchers have designed enzymes from scratch with AI; and Google launched AlphaFold, its protein-structure prediction tool, in 2020. Artificial intelligence is already embedded deep in biological research, well beyond the labs of the AI companies themselves.