AI Can Design New Proteins: Why Biosecurity Guardrails Are Now Essential

The Capability Is Real. The Guardrails Aren’t.

AI protein design tools are no longer experimental curiosities. They’re actively used across medicine, materials science, and fundamental biology research. The same models that could accelerate drug discovery can also, in the wrong hands, be used to engineer harmful biological components.

Here’s the uncomfortable truth: most users — legitimate researchers and bad actors alike — can access powerful biological AI models today without encountering a single meaningful guardrails.

Compare that to large language models. Companies developing frontier LLMs have at least begun building safeguards: responsible-use policies, threat models, red-teaming, and model evaluations. These aren’t perfect, but they exist and they’re deployed at scale.

For biological AI models, particularly open protein design tools, almost nothing equivalent is in place.

Why Sequence-Based Screening Isn’t Enough

One common biosecurity approach is DNA synthesis screening — checking whether a requested sequence resembles known harmful biological material. It’s a reasonable starting point, but it has a critical blind spot.

Biological AI models can generate novel protein designs that bear little resemblance to anything found in nature. A screening method built on sequence similarity will simply miss threats that don’t look like known threats.

That’s the fundamental problem with applying old screening logic to new AI-generated outputs.

A Different Approach: Screen Structure and Function

Work emerging from biosecurity researchers — including contributors to NTI | bio’s proposed input screening framework — points toward a more robust method: evaluate what a protein does and how it’s shaped, not just the letters in its sequence.

The approach works by screening user input sequences of protein binding targets using a protein language model’s embedding space. This computational method encodes biological meaning in a way that captures functional properties, not just surface-level sequence similarity.

There’s a practical advantage to screening inputs rather than outputs. Known protein binding targets can be systematically flagged before the AI model generates a novel design. That’s a much more tractable problem than trying to interpret a genuinely novel protein after the fact.

Guardrails and Trusted Access Work Together

Input screening alone isn’t a complete solution. The biosecurity community is increasingly pointing toward a layered approach that combines:

• Input screening tools that flag potentially dangerous binding targets before a design is generated
• Trusted user access programs that verify who is using powerful biological AI models and for what purpose
• Ongoing refinement of the underlying databases of potentially harmful targets

Trusted user access programs are already being developed by organizations like the Coalition for Epidemic Preparedness Innovations. Some AI developers have begun deploying similar frameworks for their own biology-focused models, suggesting the concept is gaining traction across the ecosystem.

The Investment Is Accelerating — Safeguards Need to Keep Pace

Governments, private companies, and research institutions are pouring resources into AIxBio capabilities. The potential benefits — faster drug development, better understanding of disease, new materials — are genuinely significant.

That investment isn’t slowing down, and it shouldn’t. But capability without accountability creates compounding risk. Every month that powerful protein design tools remain accessible without meaningful guardrails is a month where the gap between what’s possible and what’s safe grows wider.

What to Watch

If you’re tracking the AI tools ecosystem, AIxBio is one of the fastest-moving and least-governed corners of it. A few things worth paying attention to:

• Whether open protein design tools begin adopting input screening as a standard feature
• How trusted user access programs scale beyond early adopters
• Whether biosecurity frameworks developed for DNA synthesis get updated to address structure- and function-based threats

The practical takeaway: biosecurity guardrails for protein design tools aren’t a constraint on what AI can do in biology. They’re the infrastructure that determines whether the field can keep operating with public trust and institutional support. Without them, one high-profile misuse event could set back legitimate research far more than any screening tool ever would.