A Startup Wants to Fight AI Bioweapons With AI

Startup Radical Numerics has developed a model that it says can identify natural and AI-engineered pathogens without an internet connection.

By Emily Mullin • Science • November 11, 2026

The Threat Landscape

As of 2026, the convergence of artificial intelligence and synthetic biology has created a new class of biosecurity risks. Large language models and generative design tools can now propose genetic sequences and protein structures at a speed and scale that traditional screening systems were never built to handle. Experts warn that these same capabilities could be misused to engineer pathogens with enhanced transmissibility or immune evasion—raising the specter of AI-assisted bioweapons.

A New Approach: Detecting Pathogens Offline

Enter Radical Numerics, a startup that believes the best defense against AI-enabled biological threats is AI itself. The company has developed a model it claims can identify both naturally occurring and AI-engineered pathogens—and do so without an internet connection.

Why Offline Matters

The decision to run the model offline is deliberate. Many biosecurity-sensitive environments—government laboratories, field diagnostic sites, and remote public health outposts—operate with limited or untrusted connectivity. An offline-capable model reduces exposure to cyberattacks, data exfiltration, and the risk of relying on cloud infrastructure that may be compromised or unavailable during a crisis.

Distinguishing Natural From Engineered Pathogens

A central challenge in biodefense is distinguishing pathogens that evolved naturally from those designed or modified in a lab. Radical Numerics says its model is trained to detect signatures of engineering—patterns in genetic sequences and protein structures that hint at deliberate design. If validated, this capability could give public health officials and security agencies a faster way to flag suspicious outbreaks and investigate potential bioweapon incidents.

What’s Next

The company’s claims will need independent validation, particularly around false positives and false negatives in real-world settings. Still, the broader trend is clear: as AI lowers the barrier to engineering biological threats, defenders are turning to AI to keep pace. Radical Numerics’ offline model represents one early attempt to build that defense—and a reminder that in 2026, biosecurity is increasingly a computational problem.

Image: A biohazard symbol rendered in replicated pixels, contained by surrounding pixels. Credit: WIRED

via Wired AI