DeepMind watermarks AI-generated proteins to deter bioweapon misuse

Google DeepMind developed the SynthIDBio family of techniques to embed hidden watermarks in AI-designed proteins: SynthIDBio-sequence encodes a key into amino acid sequences from ProteinMPNN, while SynthIDBio-structure tweaks 3D structures via AlphaFold 3. Detectors flagged over 99% of watermarked proteins and correctly ignored 99.9% of unmarked ones, though experts warn the marks are easy to strip.

• SynthIDBio-sequence encodes a watermark in amino acids chosen by ProteinMPNN
• SynthIDBio-structure makes tiny edits to structures predicted by AlphaFold 3
• Detection caught over 99% of marked proteins and 99.9% of clean ones
• Watermarks can be removed by resequencing or adding noise to coordinates