Study Finds AI-Designed Toxin Variants Could Evade DNA Screening Tools
Summary
Researchers tested about 75,000 AI-designed variants of 72 toxins against four DNA-order screening programs. Detection varied, and three packages were updated; the study relied on predicted structures and did not establish that the variants were functional toxins.
Key points
- Researchers used three open-source AI packages to generate about 75,000 variants based on 72 toxins, including ricin.
- Four DNA-order screening programs varied widely in how well they flagged the variants as threats.
- Three screening packages were updated, improving their ability to identify variants.
- Some structurally similar variants still escaped patched screening; the article reports these as 1–3% of the “very similar” group.
- The evaluation relied on software predictions, not biological tests, so it did not establish which variants could function as toxins.
- Researchers said the approach did not appear to present a major practical threat at present, but could help screening developers prepare for advances in AI protein design.
Article Details
- Event Type
- Research disclosure of a potential biosecurity vulnerability in DNA-order screening systems
- Impact
- Researchers led by Microsoft found that AI-designed variants of toxins could evade some DNA-order screening software. The analysis evaluated about 75,000 predicted variants of 72 toxins; it did not establish that the variants were functional toxins. Three screening packages were updated, while one company judged the risk insufficient to warrant a security patch. The authors described the practical threat as limited at present, but warned that advances in AI protein design could increase future screening challenges.