Anthropic Adjusts Fable 5 Biology Safeguards the Same Day Stanford Work Shows AI Can Design Viral Genomes
What changed
Anthropic updated its Fable 5 biology safeguards on August 7, the same day researchers from Stanford and the Arc Institute published a paper in Science demonstrating that their Evo 1 and Evo 2 models can design fully functional viral genomes. The timing is not coincidental. The question Anthropic is answering is not whether AI can do biology — Stanford just proved it can — but who gets to control the terms of access. The update reduces biology-related fallbacks on benign queries by roughly 85 percent. Anthropic rewrote and retrained its safety classifier’s constitution to better distinguish everyday health and educational questions — interpreting lab results, understanding symptoms, learning about biology — from dual-use research that remains blocked. Queries touching virology, toxicology, molecular design, and drug design still get routed to Opus 5, a less capable model. Fable 5 is not yet usable for professional biology research or drug development. But the company is building the infrastructure to change that. Anthropic says it is developing trusted access pathways for frontier biology capabilities — a vetted, gated model for researchers who need the full power of Fable 5 without open-ended exposure. The company’s own capability assessments acknowledge that Fable 5 could provide significant uplift to a malicious actor — capabilities they could not find anywhere else. The contrast with Stanford’s work is the structural frame. Evo 2 is an open-weight model, available for public download. As The Guardian and CNN reported, the ability to compose viral genomes with generative AI now exists while governance does not. A companion commentary in Science from Johns Hopkins researchers warned that the study raises urgent biosafety and biosecurity questions. The US Intelligence Community 2026 Annual Threat Assessment explicitly notes that advances in biotechnology could lead to novel biological threats from state actors maintaining offensive weapons programs. Anthropic’s response — restrict the dangerous queries, open the benign ones, build a gated pathway for everything in between — maps directly onto the regulatory asymmetry the company needs. The White House AI Framework, finalized August 4, excludes open-weight models from federal security review entirely. Closed-source labs face 30-day voluntary early access review delays; open-weight developers ship with zero…
Sources and evidence
外部发现 · Anthropic · Original source ↗ · Site entry ·