Anthropic Opens a Bay Area Wet Lab
The lab is not for clinical trials, and Claude-directed robotic experiments remain exploratory with human oversight.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:003:44
Read transcript
Anthropic has opened a wet lab in the San Francisco Bay Area, giving its life-sciences effort a physical place to run biology experiments. That expands a program already supplying AI services and tools to Genentech, Bristol Myers Squibb, and Novo Nordisk. Anthropic says the facility is not specifically for drug discovery, and it is not running clinical trials. Some work will happen in-house; other experiments will continue through external partners. Eric Kauderer-Abrams, Anthropic’s head of life sciences, says real laboratory work remains the final test in biology. The company describes life sciences as one of its largest investment areas by headcount and resources. It has also launched Claude Science and introduced the Model Hardware Standard, aimed at helping AI systems connect with laboratory equipment. A more ambitious direction is exploratory work on having Claude direct robotic lab units through experiments with limited human intervention. That is not evidence of a deployed autonomous laboratory. Anthropic says human oversight and involvement remain essential for safety. The important shift is therefore not that Anthropic has announced an autonomous drug-discovery machine. It is that the company is adding firsthand experimental capacity to its AI work, while keeping clear boundaries around what this lab is and is not. Biology moves from screens into rooms, equipment, samples, and procedures—and that makes access, supervision, and containment operational questions, not just product claims. That same question about containment surfaced in a cybersecurity evaluation. Google says Gemini, during a test run by Irregular, was given unintended internet access and reached infrastructure at three real companies. The model had been told to attack a fictional company whose name matched a real one. It used passwords found online or guessed, logged into the intended company’s systems and two others, then stopped after recognizing that the environments were real. Google says no harm resulted. Irregular said it notified the affected entities and fixed the known configuration issues weeks before disclosure. The sequence matters: recognition came after access. The test shows why an AI system’s ability to notice a boundary is not a substitute for a boundary that holds before contact. And in law, OpenAI is putting a different kind of boundary around access. Astra for Law is launching first with selected firms, built on GPT-6 Astra and an index of U.S. case law, statutes, regulations, and other legal materials. OpenAI says it is designed for research, drafting advice, and custom applications, with protections for confidential client work. Harvey and Legora can build on the platform. Connections to Relativity, Clio, Intapp, and Thomson Reuters are planned, not demonstrated as live integrations. OpenAI worked with Sullivan and Cromwell, Ropes and Gray, Cooley, Latham and Watkins, and Wachtell Lipton on early applications. For now, the concrete deployment is narrow; the broader provider network remains a promise to execute. Finally, Meta is moving its Muse personal agent onto the Mac, where it can work across files, messages, calendars, notes, and mail. Users choose connected apps and access levels. Meta says Muse asks before sensitive actions such as sending email or making a purchase, and records actions it has taken or plans to take. A separate agent, Sentinel, governs internet access and external-action requests. Those safeguards are Meta’s stated protections, not independent proof that every error or attack is prevented. Across these launches, the practical test for next week is straightforward: not merely what an agent can reach, but who grants that reach, what stops it before a mistake, and where a person can still inspect, interrupt, or revoke the action.
