Anthropic Quietly Ran a Biology Lab for Months. Its First Claim Is Already Disputed.
Bot Mutiny |
Anthropic says roughly 950 Claude agents found a new enzyme system with CRISPR-like repeats. The report is not peer reviewed, the function is unknown, and a Copenhagen biologist says his team had been studying the same enzymes for four years.
On September 23, Anthropic announced that Claude had found a new enzyme system in the DNA of viruses that infect bacteria. The company says Claude made the find "with only high-level direction from our scientists." It calls the system array-associated reverse transcriptases, or ART, and compares its layout to CRISPR.
The announcement also introduced something the public had not been told about until days earlier. Anthropic runs its own biology lab.
A lab that was already running
TechCrunch reported on September 18 that Anthropic "has confirmed to TechCrunch" it operates a wet biology lab in the Bay Area. Anthropic declined to give TechCrunch specifics on the lab's work. TechCrunch also reported that the company would not say how many months the lab had been open.
The Scientist reported that the company said it "had been quietly operating the lab since the spring of this year." Anthropic's own post says the research group formed in spring 2026.
What Anthropic discloses about oversight
The company's post gives a short account of safety conditions. It says the lab works only at biosafety levels BSL-1 and BSL-2 and does "not handle pathogens that can infect humans." It says "all of the lab work is performed by human scientists."
The post describes no outside oversight of the lab itself, such as an institutional biosafety committee or an external audit. The technical report lists six authors, all with an Anthropic affiliation.
TechCrunch placed the lab in context. It noted that Dario Amodei, Anthropic's chief executive, has said one of the things he fears most is AI being used for bioterrorism. Amodei, as quoted by TechCrunch, said Claude may one day run lab equipment itself, "but we aren't doing that today."
How much of it was Claude
Anthropic's post says its scientists' "involvement was limited to the initial prompt and the lab work." Amodei put it differently, according to TechCrunch. He wrote that the system was found "mostly, though not entirely, by Claude."
The technical report describes two phases. In the first, the search campaign ran "without human intervention" and ended with written reports delivered to human reviewers. In the second, the report says "the authors directed the analysis and Claude wrote and ran the code."
The numbers also shift between documents. The blog post cites roughly 950 agents, 210 million tokens and 21 hours. The technical report counts 949 agent sessions, 215.6 million tokens and 21.5 hours of wall-clock time.
Not reviewed, function unknown
Anthropic calls its write-up a pre-print. The Scientist describes it as a companion technical report "which has not been peer-reviewed."
Anthropic does not know what ART does. Its post says work to understand the system's primary function "is ongoing." The technical report is more specific. It says the authors "have not shown that the RT is active." It adds that what the system does for the phage is "currently unknown." Amodei, as quoted by ThePrint, wrote that its "biotechnological utility (if any)" is not yet clear.
What outside scientists said
Anthropic's post quotes Feng Zhang of MIT and the Broad Institute, who reviewed the pre-print. Zhang called the finding "genuinely intriguing" and said it "merits further investigation."
Other researchers questioned the claim of discovery. The New York Times, in a story by Carl Zimmer republished by the Irish Times, reported that "some scientists were quick to cast doubt on the achievement."
Mario RodrÃguez Mestre, a computational biologist at the University of Copenhagen, said he and his colleagues had been studying the same enzymes for four years. The Times reported that Mestre first encountered the enzymes in jumbo phages in 2022. His team had nicknamed them "jumbotrons." The team has not published its findings. The Times reported that the team regularly used Anthropic's models to write code and draft manuscripts, and that they shared unpublished information with Claude.
Mestre told the Times his concern is that "this information was used to train future versions of the models." He also said: "we were aware of these repeats more than one year ago."
Seth Shipman, a bioengineer at the Gladstone Institutes and UCSF, told the Times: "Many of us know about these systems."
Anthropic responded in a statement. It said it is "not aware of any previously published work describing the ART system." It said Claude "was also not trained on any user transcripts." It added that its molecular biology team has no such access.
Anthropic's own post concedes that the underlying reverse transcriptase "had been identified in previous studies." TechCrunch reported that Amodei acknowledged a Stanford team had earlier found a system "in some ways similar to the one Claude found."
What this does not establish
None of the reporting shows that Anthropic's lab has caused harm or broken a rule. Anthropic says it works at low biosafety levels with no human pathogens. No source cited here disputes that.
The reporting does not settle who found ART first or how Claude came to flag it. Mestre has raised a question. Anthropic has denied training on user transcripts. No source cited here presents evidence that resolves it.
The technical report has not been peer reviewed. ART's function is unknown, and Anthropic's comparison to CRISPR rests on the layout of the system, not on any shown activity. What the public knows about how the lab is overseen comes from Anthropic's own short description.