Anthropic Spent Months Warning the World About AI Bioweapons While Quietly Building a Biology Lab
The company whose CEO says AI-assisted bioterrorism could arrive within a year has, since spring, been running its own wet lab where Claude designs biology experiments that human scientists then physically carry out.
Since spring 2026, Anthropic has operated a molecular biology wet lab somewhere in the Bay Area — real benches, real reagents, real human scientists in real coats, executing real experiments that Claude designs. The company confirmed the lab exists only after being asked about it, and even then declined to say where it is, how big it is, how many people work there, or what it cost. What it did confirm is more interesting than any of that: the lab produced a genuinely novel result, an enzyme system called array-associated reverse transcriptases, which Claude apparently identified and which Anthropic says can manipulate DNA. Dario Amodei, in the same breath as the announcement, noted that Stanford had “previously discovered a system that is in some ways similar to the one Claude found” — an asterisk on the victory lap that Anthropic included itself, which counts for something.
The lab operates at biosafety levels 1 and 2, the two lowest tiers, with no pathogens capable of infecting humans. Claude does not touch equipment; it proposes experiments and human scientists run them. In one enzyme-discovery push, the company says, it burned through roughly 950 agents, 210 million tokens, and 21 hours of compute to get there — an oddly specific accounting for a company that won’t say what the building costs to keep the lights on.
Here is the part that requires a calendar. One week before this lab became public knowledge — via Reuters, not via Anthropic — the company had published a report about shutting down Claude accounts it believed were being used to help develop biological weapons, including work on making avian flu more transmissible in mammals. Around the same time, Amodei was touring the essay circuit warning that AI-assisted bioterrorism carried a “substantial risk” of arriving within two to three years, an argument he has made central to his broader case that the entire industry needs to slow down. At no point in any of that did he mention that his own company had, some months earlier, opened a physical facility for running real biological experiments designed by the exact kind of model he was warning about.
Anthropic’s position, when reporters caught up with it, was that the lab is “not for drug discovery specifically,” an answer that manages to rule out one thing without confirming anything else. Eric Kauderer-Abrams, the company’s head of life sciences, offered the closest thing to an explanation on record: “We believe that to do biology, the final test is still and will be for a while in real lab work. We absolutely are doing that today.” Asked about the obvious tension between running biology experiments in-house and warning the public that biology experiments are exactly how a bad actor builds a weapon, Kauderer-Abrams allowed that “we’re always balancing these things,” which is the kind of sentence that sounds like an answer until you write it down.
None of this happened by accident, budget-wise. Life sciences is reportedly now one of Anthropic’s largest areas of spending, on top of the roughly $400 million acquisition of Coefficient Bio and the launch of Claude Science, a dedicated product for laboratory research whose customer list already includes Novo Nordisk, Genentech, and Bristol Myers Squibb. This is not a side project run by a curious research team with a spare budget line. It is a company simultaneously building the commercial pharma business it says it isn’t building and warning regulators that the underlying capability is dangerous enough to require industry-wide restraint — the same week it’s recruiting a former “no FDA for AI” White House adviser to help design what that restraint should look like.
There is a version of this story where all of it is fine, even admirable: a company takes its own warnings seriously enough to build the verification infrastructure needed to actually test what its models can do in a lab, rather than just speculating about it from a blog post. Owning the wet lab means Anthropic doesn’t have to rely on an outside partner to confirm or deny whether Claude actually found something real, which is precisely how it caught its own hedge about the Stanford overlap instead of getting caught by someone else. That is a more responsible posture than most of the industry manages. It is also, inescapably, the same company that spent the run-up to a wet lab’s public debut telling everyone how frightened it was of exactly what a wet lab could do.
Anthropic has not disclosed a biosafety officer, an external audit, or a public dual-use review process for what the lab is allowed to work on next — only that the current outputs are pharmaceutical-adjacent and that, per Amodei, the company would “stop before clinical trials” to avoid competing directly with the drug companies who are also its customers. Whether that’s a firm internal line or a marketing distinction that dissolves the first time a compound looks promising enough is, like the lab’s square footage, not something Anthropic has decided to share yet. The essay said slow down. The lab, apparently, had already sped up.
Sources: