Microsoft-Led Study Reveals AI-Crafted Toxins Evade DNA Screens

Microsoft has uncovered a significant vulnerability in DNA biosecurity systems, revealing that AI-generated toxins can evade detection. In a study published in Science, researchers demonstrated that AI tools could redesign toxic proteins, such as ricin, in ways that bypass existing DNA screening protocols.

The study, led by Microsoft's Eric Horvitz and Bruce Wittmann, involved creating over 70,000 synthetic DNA sequences of controlled toxins. These sequences were tested against biosecurity screening software used by DNA synthesis companies. The results showed that while native toxin sequences were often flagged, many AI-altered variants went undetected, with some detection rates dropping to zero.

Following the study, most DNA screening vendors issued software updates, improving detection rates to an average of 72%. However, experts warn that these updates are only temporary solutions. The study highlights the need for integrating safeguards directly into AI protein design tools to prevent misuse and ensure biosecurity.