Anthropic has confirmed that it operates a physical "wet" biology laboratory in the San Francisco Bay Area, where AI models direct robotic equipment to conduct real-world experiments with minimal human intervention. The lab's existence was first reported by Reuters through anonymous sources. Anthropic subsequently confirmed it to TechCrunch.

The confirmation came the same week Anthropic published a report documenting blocked attempts to misuse Claude for biological weapons research. The timing was not lost on observers.

The lab focuses on protein design research conducted both computationally ("in silico") and physically ("in vitro"). Kauderer-Abrams characterized the operations as consistent with standard biotech company practices. Anthropic has stated it will not pursue clinical trials, positioning its work before the point where it would compete directly with pharmaceutical partners and customers.

In August 2026, Anthropic published results from its Claude Science platform (launched June 30, 2026) showing that Claude-designed protein binders succeeded against 14 of 15 targets tested. Two independent firms, Adaptyv Bio and Twist Bioscience, built and validated the designs in wet labs.

The workflow used a single expert-written prompt of approximately 30,000 tokens. Claude orchestrated existing open-source tools including RFdiffusion and ProteinMPNN. The system used Claude Mythos Preview and Claude Opus 4.8 for protein design, with up to 12,500 NVIDIA H100 GPU-hours for multi-target campaigns.

The program offers two tiers. Standard Use provides team-wide access for routine life sciences research, supply chain, and manufacturing work, renewing annually. High-risk Use grants project-specific access for work that standard safeguards would block, limited to a single named research project with six-month renewals.

Applicant organizations are evaluated on research credentials, security standards, and ethical oversight. The program replaces real-time biology query blocking with offline monitoring and customer triage. Traffic requires 30-day retention for safety analysis but cannot be used for model training. PHI must be kept in a separate HIPAA environment and cannot enter LSVP workspaces.

The lab announcement arrived in a week when Anthropic was simultaneously publicizing biosecurity risks from AI. Days before the Reuters report, Anthropic published a report documenting blocked misuse of Claude between December 2025 and August 2026. One case involved actors seeking help with a grant application for gain-of-function research on chikungunya virus, aiming to enhance its transmissibility and immune evasion.

Separately, Anthropic CEO Dario Amodei has previously described AI-assisted "bioterrorism" as a leading risk, and the company's own safety researcher Evan Hubinger has publicly estimated a greater than 10% probability that AI "could kill all humans" within a decade. A former researcher told the BBC that staff were "genuinely frightened" about humanity's future.

Investor Chamath Palihapitiya drew the comparison publicly, likening Anthropic's position to Philip Morris acknowledging tobacco's harms while continuing to sell cigarettes. Co-host David Sacks argued the company must either formally repudiate its safety researchers' warnings or formally acknowledge them, either outcome potentially complicating the company's rumored $2.3 trillion IPO.

MIT synthetic biologist Kevin Esvelt has reported that an AI chatbot disclosed "a novel form of bioweapon" he had not previously considered, though he noted that Claude's safety classifier did trigger correctly on a bioweapon-related post he made. He has argued publicly that researchers who dismiss civilization-level bioweapon risk may be wrong, since "you can't see a way, or evidence that others can, doesn't mean no one could."

Stanford biomedical ethics professor David Magnus called the broader situation "very scary." University of Hamburg biosecurity researcher Dunja Sabra assessed that "the chances are that someone determined would succeed eventually." Cornell computer scientist John Thickstun highlighted the governance gap, noting that AI companies are making "value judgments at societal scale without any kind of democratic or deliberative oversight."

Some Imperial College London biologists have countered that current AI tools remain insufficient for full bioweapon development and that existing safeguards are adequate.