A biological 0-day? Threat-screening tools may miss AI-designed proteins. (arstechnica.com)

A Microsoft-led research team announced they had found—and in some cases patched—a "biological zero-day" in commercial DNA synthesis screening: AI-designed proteins can produce novel DNA sequences that encode toxic or dangerous proteins yet evade current filters. Today, DNA-order screening relies largely on sequence similarity and known-protein matching (including accounting for synonymous DNA variants) to flag suspicious orders. The researchers showed that modern protein-design algorithms can generate de novo proteins with toxic function or active sites that are functionally similar but sequence-divergent from known threats, allowing them to slip past those similarity-based checks. This matters because it shifts the attack surface from recognizable sequences to emergent functionality that conventional pipelines weren’t built to detect. Technical remedies include moving from pure sequence matching to function-aware approaches—structure and active-site prediction, motif detection, protein-level similarity metrics, and ML classifiers trained to recognize biochemical activities—plus updated threat models, benchmarked red-teaming, and industry coordination to balance sensitivity, false positives, and privacy. The finding underscores an urgent need for cross-disciplinary collaboration between AI protein designers, biosecurity experts, and DNA synthesis providers to revise screening standards before more sophisticated or adversarial designs become widespread.