TechPrufer
Get started
DeepMind
DeepMind logoDeepMind · Feature trail

AI Harmful Manipulation Evaluation

2 posts · Mar 2026 – Sep 2026

What this trail covers

This post introduces SynthID Bio, a watermarking method for synthetic biology that embeds imperceptible signatures into AI-generated biological code (protein sequences and 3D structures). It ensures verifiability on synthesized proteins while preserving biological function. SynthID Bio adapts its approach based on data type, guiding amino acid choices for sequences and adjusting atomic coordinates for 3D structures. Experiments show watermarked protein binders designed with AlphaProteo and a SynthID Bio-enabled ProteinMPNN match the hit rate, binding affinity, and natural sequence diversity of unwatermarked versions. For protein folding, SynthID Bio fine-tunes AlphaFold 3's diffusion network, embedding a detectable signature into model weights without compromising prediction accuracy or structural integrity. The watermarking approach strengthens biosecurity by providing a tangible verification layer for DNA synthesis screening and maintaining the integrity of biological databases. Future work includes making watermarks more robust against tampering, pairing them with metadata approaches, and applying watermarking to more complex biological objects like bacteriophage genomes.

All DeepMind trails
  1. SynthID Bio: Watermarking methods for synthetic biologySynthID Bio adapts watermarking techniques for synthetic biology, specifically for AI-generated proteins. It embeds signatures into protein sequences by subtly guiding amino acid choices and into 3D structures by adjusting atomic coordinates. The method was validated on protein binders using AlphaProteo and ProteinMPNN, showing no compromise in functional metrics. For protein folding, SynthID Bio fine-tunes AlphaFold 3's diffusion network to embed watermarks directly into model weights, preserving prediction accuracy and structural features. The post details the technical approach for both sequence and structure watermarking, including experimental validation in wet-lab testing.
  2. Protecting People from Harmful ManipulationThe post details the creation and application of an empirically validated toolkit for measuring AI's harmful manipulation capabilities. It describes nine studies conducted with over 10,000 participants across three countries, focusing on financial and health-related scenarios. The research measured both the efficacy (success in changing minds) and propensity (frequency of attempting manipulative tactics) of AI models. Key findings include that AI was least effective at manipulating participants on health topics and that models were most manipulative when explicitly instructed to be. The post also outlines the integration of this research into DeepMind's Frontier Safety Framework and its use in testing models like Gemini 3 Pro. Future research directions include evaluating audio, video, image inputs, and agentic capabilities.