October 2, 2025, NPR (National Public Radio) reported a high‑profile study: researchers used generative AI to "brainstorm" dangerous proteins or DNA fragments. These AI‑generated sequences showed potential hazardous functionality in predictive tests but were able to evade the screening systems widely used by DNA synthesis companies.
Most DNA synthesis vendors currently rely on sequence alignment and databases of known hazardous genes to intercept suspicious orders, for example using BLAST to detect homology with threat factors or pathogen genes. The study notes that AI can design sequences that do not directly match known hazardous genes by recoding codons, dispersing pathogenic function across longer sequences, or embedding "apparently harmless" segments, thereby lowering alignment scores and evading alignment‑based screens.
The research team ran AI‑generated sequences through simulated existing screening workflows and found many designs predicted to have specific biological activity did not trigger safety thresholds — meaning they could technically form dangerous proteins or interfere with biological processes, yet would not be blocked under current detection rules. The authors stress this is a "conceptual" demonstration intended to highlight potential blind spots in defenses, not evidence of an existing malicious event.
The study outlines several classes of potentially misusable techniques: first, codon replacement (synonymous mutations) that alters sequence appearance while preserving protein function; second, splitting hazardous functions into multiple fragments that appear harmless individually but are harmful when combined; third, using AI to design novel protein domains that achieve specific effects structurally without sequence similarity. The researchers warn that combined with open‑source or commercial generative models, these techniques push biosecurity defenses into new challenges.
Recommended responses include: DNA synthesis companies moving beyond simple sequence alignment toward function‑prediction‑based screening tools; industry groups such as the International Gene Synthesis Consortium (IGSC) updating rules and strengthening information sharing; and AI model developers working with the biotech sector to set safe‑use guidelines and explore embedding "abuse mitigation" measures or watermarking in models. The team also calls on regulators and funders to increase investment in this cross‑disciplinary risk and to advance detection and review capabilities.
The report notes the study did not publish directly reproducible "harmful recipes" and that the researchers followed responsible disclosure practices, but the findings have already reverberated through academia and industry. For early‑career scientists and entrepreneurs, the result is both a warning and a reminder: AI accelerates biological design innovation while reshaping risk, and biosecurity and AI governance must advance together to keep innovation and risk within acceptable bounds.