Researchers at Microsoft have uncovered a serious weakness in global biosecurity controls by showing how artificial intelligence (AI) can evade detection systems designed to stop dangerous synthetic proteins from being created. The study revealed that AI tools can modify the genetic code of lethal toxins in a way that keeps their harmful structure intact but “rewrites” their DNA sequence to slip past standard safety screening used by DNA synthesis companies.
The team, led by Microsoft Chief Scientific Officer Eric Horvitz, performed a large-scale digital experiment generating over 75,000 variants of toxic proteins, including well-known biological threats like ricin and botulinum toxin. These AI-altered sequences largely escaped the usual screening software, raising concerns about the ability of bad actors to misuse AI for bioweapon development.
Following this discovery, Microsoft worked with Twist Bioscience and other industry players in a 10-month collaborative effort to improve biosecurity software. Updates to the screening systems now detect roughly 97% of the AI-generated variants, though a small percentage still bypasses checks.
This vulnerability is linked to the same AI technology driving advances in drug discovery, which uses protein design algorithms to create new molecules for medical purposes. While these studies were purely digital and did not produce actual toxins, they highlight the growing dual-use risk of AI in biology, where powerful tools can be used for both good and harm.
Emily Leproust, CEO of Twist Bioscience, underscored the need for constant vigilance and rapid adaptation of screening practices to keep pace with AI’s evolving capabilities. DNA synthesis companies generally refer very few suspicious orders to authorities, showing the high level of trust placed in these biosecurity measures.
