Anthropic's Claude AI discovers new gene-editing system

Anthropic has revealed that its proprietary AI model, Claude, successfully identified a previously unknown enzyme system capable of performing gene-editing functions. Operating within the company's newly established wet biology lab in the San Francisco Bay Area, the AI analyzed genetic data from bacteriophages to locate the system, which can cut, copy, and paste DNA. While human scientists performed the physical validation experiments, Amodei noted the discovery was made "mostly, though not entirely, by Claude" during computational phases.

The speed of the discovery highlights the efficiency of agentic AI workflows. While the physical laboratory was only established this spring, Claude completed the computational search in just 21 hours of active processing. To achieve this, the system deployed approximately 950 autonomous agents that processed 210 million tokens. This rapid turnaround showcases how large language models can drastically compress the early stages of biological research, which typically require months or years of manual human analysis.

Despite the highly automated nature of the search, Anthropic is maintaining strict safety boundaries. The Bay Area facility operates under Biosafety Level 1 (BSL-1) and Biosafety Level 2 (BSL-2) protocols, meaning researchers do not handle pathogens capable of infecting humans. Anthropic CEO Dario Amodei, who has voiced concerns over AI-enabled bioterrorism, believes AI could help cure most diseases in 5 to 10 years. He emphasized that human scientists still execute all physical lab work, though he noted that Claude could eventually control laboratory equipment autonomously once appropriate safeguards are established.

For biotechnology practitioners, this milestone signals a shift toward highly automated pipeline discovery. While other organizations have leveraged machine learning for biology—such as Google launching AlphaFold in 2020 or researchers at UC San Francisco designing synthetic enzymes—Anthropic's agent-driven approach suggests that LLMs can actively propose and locate novel biological mechanisms. Although the broader scientific community must still validate the findings, the integration of massive agent networks could soon make AI-driven discovery a standard starting point for genetic engineering.