The Paraphrase Project investigated how open-source AI protein design tools could be exploited to generate toxic proteins that evade existing DNA synthesis screening systems. Conducted over two years by a cross-sector team including Microsoft, Twist Bioscience, and IBBIS, the work demonstrated that toxic proteins such as ricin can be paraphrased by AI models to maintain structure and function while avoiding detection. This vulnerability was described as the first “zero day” in AI biosecurity, prompting a cybersecurity-style emergency response team approach.
The team developed patches to existing screening software that significantly improved detection of AI-rewritten toxic proteins across multiple companies. These efforts expanded to cover 72 hazardous proteins including toxins and viral proteins, with results validated by predicting the structure and function of paraphrased sequences. Members of the International Gene Synthesis Consortium appreciated the coordinated disclosure and mitigation measures, which now form part of an ongoing biosecurity framework.
Addressing the risks of information hazards, the project designed a novel system to tier and gatekeep sensitive data and software, balancing openness with harm prevention. This approach offers a model for managing biosecurity risks from dual-use technologies while fostering collaboration. The researchers emphasize that AI biosecurity remains an evolving challenge requiring layered defenses at both the technical and policy levels to anticipate future advances in protein design capabilities.
👉 Pročitaj original: Microsoft Research AI
