TL;DR: Anthropic reports that 17 AI-assisted protein designs bound to scorpion toxins in lab tests. Whether they can neutralize the toxins is still untested.

An internal Claude research model coordinated existing protein-design software. Lab testing of 211 candidates produced the 17 hits, targeting four toxins; the team is releasing the designs and measurements.

The useful result is a design workflow checked against physical experiments. Binding is an early step: the work has not established an effective antivenom or safety in animals or people.