Google DeepMind unveils hidden watermark to flag AI-crafted proteins

Google DeepMind scientists have developed SynthIDBio, a method that subtly encodes a hidden marker into the sequence and shape of proteins created by artificial-intelligence tools such as AlphaFold. The watermark is designed to be invisible to functional performance while allowing trusted parties to detect AI-originated proteins using a secret key. Biosecurity scholars say the approach could safeguard research databases and screening processes, though it can be stripped away by re-running the design through another algorithm.

The technique was stress-tested on proteins involved in viral entry, blood-vessel growth and immune regulation, showing no loss of activity. DNA-synthesis providers could adopt the system to flag novel AI-designed proteins that might evade traditional toxicity checks. While not a complete solution, the watermark adds another layer to bio-security defenses in a rapidly evolving field.