Paraphrase Project

The Paraphrase Project addresses a critical and emerging risk: as generative AI and synthetic biology advance, there is a growing possibility that open-source AI tools could be misused to design biological toxins that evade current biosecurity screening systems. The project’s core motivation is to preemptively identify and patch vulnerabilities in biosecurity protocols before they can be exploited, ensuring that the rapid pace of innovation in AI-driven biotechnology does not outstrip safety measures. This work is especially urgent as DNA synthesis and protein engineering become more accessible globally, raising the stakes for both accidental and deliberate misuse.

This cross-sector initiative:

• Demonstrates how AI models can “paraphrase” proteins—rewriting amino acid sequences to preserve biological function while altering their form, potentially allowing dangerous sequences to slip past existing biosecurity filters.
• Tests the limits of current DNA synthesis screening tools by generating thousands of synthetic variants of known toxins (like ricin) and evaluating whether these variants are detected.
• Develops and validates new detection algorithms and protocols for red-teaming biosecurity systems, establishing a framework for managing information hazards and responsible disclosure in sensitive research

Paraphrase Project contributors: Bruce J. Wittmann (Microsoft), Tessa Alexanian (The International Biosecurity and Biosafety Initiative for Science/IBBIS), Craig Bartling (Battelle), Jacob Beal (RTX BBN), Adam Clore (Integrated DNA Technologies Inc/IDT), James Diggans (Twist Bioscience), Kevin Flyangolts (Aclid), Bryan T. Gemler (Battelle), Tom Mitchell (RTX BBN), Steven T. Murphy (RTX BBN), Nicole E. Wheeler University of Birmingham), Eric Horvitz (Microsoft)