Back
Our approach to biosecurity for AlphaFold 3
webstorage.googleapis.com·storage.googleapis.com/deepmind-media/DeepMind.com/Blog/a...
This is DeepMind's official biosecurity policy document accompanying the AlphaFold 3 release, relevant to discussions of dual-use AI governance and responsible deployment practices for high-capability biological AI systems.
Metadata
Importance: 62/100blog postprimary source
Summary
DeepMind outlines the biosecurity measures and risk mitigation strategies implemented for AlphaFold 3, addressing concerns about dual-use potential of a powerful protein structure prediction system. The document explains how DeepMind assessed misuse risks and what safeguards were put in place before releasing the model, serving as a case study in responsible deployment of dual-use AI capabilities.
Key Points
- •DeepMind conducted biosecurity risk assessments prior to releasing AlphaFold 3, consulting external experts on potential misuse scenarios.
- •Access controls and usage restrictions were implemented to limit the model's accessibility for harmful applications, particularly in bioweapons contexts.
- •The document exemplifies how AI labs can attempt to balance open scientific benefit against dual-use risks when releasing powerful biological AI tools.
- •AlphaFold 3 expands beyond proteins to predict structures of DNA, RNA, and small molecules, increasing both utility and potential misuse surface.
- •DeepMind's approach reflects broader industry debates about responsible disclosure and staged access for AI systems with biosecurity implications.
Cited by 1 page
| Page | Type | Quality |
|---|---|---|
| Bioweapons Risk | Risk | 91.0 |
Cached Content Preview
HTTP 200Fetched Mar 20, 202621 KB
# Our approach to biosecurity for AlphaFold 3
**Authors: Conor Griffin, Heidi Howard, Antonia Paterson, Nick Swanson, Dawn Bloxwich,** **John Jumper, Pushmeet Kohli, Nicklas Lundblad**
### We believe that AlphaFold 3 will help to accelerate progress on tackling
### priority biosecurity challenges, and that it can be safely released. Below,
we explain how we think about biosecurity, how we approached it for AlphaFold 3, and our plans for future work.
Today we announced AlphaFold 3, a revolutionary model which can predict the structures and interactions of all life's molecules with state-of-the-art accuracy, providing valuable insights into how proteins interact with DNA, RNA and other biomolecules. Scientists can access the majority of its capabilities, for free, through our newly launched AlphaFold Server, an easy-to-use research tool.
We began exploring the potential effects of our research on biosecurity in the early days of AlphaFold, in 2020. Back then, discussions about ‘AI and biosecurity’ were rare and often limited to a handful of individuals. Today the topic has become much more prominent given the recent acceleration of progress in AI and biology. Last year, biosecurity was mentioned in both the White House Commitments and the US Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. The UK & US AI Safety Institutes and other government bodies are also considering how to best assess AI biosecurity risks, while civil society actors, such as the Nuclear Threat Initiative (NTI), are establishing dedicated forums to identify and share AI and biosecurity best practices. At Google DeepMind, we actively contribute to and participate in these deliberations.
Based on consultations with a broad community of experts, and early evidence on how AlphaFold 2 is being used, we believe that AlphaFold 3 will help researchers make progress **on priority biosecurity challenges, such as pandemic preparedness and neglected** **tropical diseases. We have also concluded that AlphaFold 3 can be safely released and does** not lead to a material uplift in risk when compared to other readily available state-of-the-art structure prediction tools.
To contribute towards ongoing AI and biosecurity discussions, we believe it is important to fi outline how we reached this conclusion. In this blog, we rst explain how we define biosecurity,
and how we think AI-more broadly-could affect it. We then walk through how we assessed the potential impact of AlphaFold 3. Finally, we conclude with some future work priorities for ourselves and the broader AI and biosecurity community. This includes a new exploration into how biosecurity screening processes could work. We will trial this for the launch of AlphaFold Server, in a narrow, targeted manner to avoid impinging on beneficial use cases. Our goal with this exploration is to inform potential screening systems for future, more powerful, AI models and tools.
## What do we me
... (truncated, 21 KB total)Resource ID:
81aa1be41165df66 | Stable ID: ZGNlNDIxNT