Last week the story was the room. The company most vocal about AI and bioterror had quietly built a wet lab in the Bay Area, and no one would say what biosafety level it ran at. This week the room produced something, and Anthropic answered the question.

The answer is BSL-1 and BSL-2. No pathogens that infect humans. Human scientists do all of the bench work. That sentence was missing from the Reuters story.

Anthropic pointed a swarm of Claude agents at a metagenomic database of about 1.9 billion protein clusters and told them to look for “interesting” reverse transcriptases (Anthropic’s word). These are enzymes that copy RNA back into DNA. Bacteria and their viruses use them in odd ways, mostly to fight each other, and several labs have been mining them for years. The last time someone combed bacterial immune systems this carefully, the result (CRISPR) won a Nobel Prize.

The agents gathered roughly 200,000 of them, sorted them, scored about 3,500 candidate partner genes, and filed a handful of reports for humans to read. In one report, an agent reading raw DNA next to a reverse transcriptase in a bacteriophage noticed a stretch of evenly spaced repeats.

Anthropic calls the system ART, for array-associated reverse transcriptases. It has three parts. An enzyme, a partner gene of unknown purpose, and an array of 3 to 21 short repeats. It looks a little like a CRISPR array. It has none of the cas genes that make CRISPR work.

Six authors, all at Anthropic, including its head of life sciences (Eric Kauderer-Abrams) and an interpretability researcher (Emmanuel Ameisen). The paper cites the prior art (retrons, diversity generating retroelements, the recent UG27 arrays).

This is a product demonstration written by the product's maker, posted on the maker's own servers rather than bioRxiv, a week after the lab became news and ten days after the Novo Nordisk deal. The outside voice is Feng Zhang, whose quote was supplied with the announcement. The paper does list what it did not show, and it reports its own failed reruns.

BSL-1 is the teaching lab (benign microbes, open bench, wash your hands). BSL-2 adds a biosafety cabinet, restricted access, and agents of moderate risk. BSL-3 is where airborne human pathogens live, behind negative pressure. BSL-4 is the moon suit.

Anthropic says it lives in the bottom two floors and handles nothing that infects humans. The ART work fits that description. Cloning phage genes into lab E. coli is about as routine as molecular biology gets. Bacteria leave on shoes, and nobody should pretend a software company is running a BSL-4.

Last week's reporting described AI models directing robots with minimal human intervention. This week Anthropic says all lab work is performed by human scientists. Both can be true (humans can supervise robots), but the emphasis has shifted.

Anthropic's leadership has spent years asking governments to regulate frontier AI rather than trust companies to grade their own homework. On wet biology, the grading is voluntary and the homework is theirs. I believe them about BSL-2. I would still prefer to believe an inspector.

The slope is the other worry. Today the agents read sequences and humans hold the pipettes. The stated ambition of Claude Science is autonomous workflows. Each step from here (agent proposes, agent designs the construct, robot runs it) is small and defensible. The sum is a different kind of lab. A BSL-2 phage enzyme lab is not gain of function on a human virus. It is also not nothing.