Mountain ViewGoogle DeepMind’s watermarked AI proteins still bound their targets.
Screening software at DNA makers cannot assess AI-designed proteins that resemble nothing known.
A secret key nudges each amino acid choice in ProteinMPNN, and only functional picks stay, leaving a statistical signature.
Very short proteins carry too few watermark amino acids, and fusing a natural protein can dilute the signal.
Its security rests on how keys from trusted universities and biotech firms get distributed and protected.
How each outlet framed it
- Ars Technica
- explains AI-designed proteins lack characterization for existing threat-detection software; Google's watermarking solution based on SynthID allows trusted research identification while opening all others to scrutiny
Sources: Ars Technica