We develop possible approaches to lower the risk of AI-assisted and AI-initiated biology. Our work falls into six areas. Three of them produce measures that can be used now, and three prepare instruments and understanding to be held against a future need.
Countermeasures placed in software
Strings of text and code that developers can add to biological and chemical software so that AI agents are deterred from operating it.
Observing and identifying agents
Honeypots and license plate readers that detect AI agents on the internet and in software environments, characterize them, and recognize them when they appear again.
Messages to the Future
Authors can opt to attach dual-use warning notices to preprints describing their software, methods, or results, to influence models during pre-training and again during inference when the models read the papers again.
Protecting research and development infrastructure
Work to find and repair vulnerabilities in the computational services on which biological and chemical research, development, and manufacturing depend, and on which emergency response may depend.
Contingency planning and emergency controls
A strict regulatory regime, published and analyzed in advance, that could be imposed on biological research, development, manufacturing and agriculture after a triggering incident.
AI and chemistry
The harms artificial intelligence might cause through chemistry, the chemical routes by which synthetic DNA is made, and the points at which regulation would operate.