The Biological Arms Race: Why Defense Must Scale at the Frontier

The core idea here is that biological intelligence--the ability for AI to read, write, and reason about the fundamental code of life--is no longer a theoretical pursuit. It is an active, accelerating arms race. The implication is that the models enabling breakthroughs in human health are inherently dual-use; they possess the same generative capabilities that could be exploited for harm. Consequently, the defensive strategy of relying on simple sequence matching is obsolete. The advantage lies with those who build defensive capabilities directly into the frontier models themselves, rather than treating them as an afterthought. For researchers and stakeholders, the divide between frontier AI and applied biology is collapsing. Those who operate at this intersection gain the ability to shape the future of both discovery and security.

The Hidden Cost of Fast Solutions

The conventional approach to biosecurity relies on sequence matching, which means comparing new DNA against a database of known pathogens. Eric Nguyen, CEO of Radical Numerics, argues this is broken because it ignores the functional manifold of biology. Just as a language model can paraphrase a sentence while keeping the meaning intact, a generative model can design a sequence that is genetically distinct but functionally identical to a pathogen.

> "A model that is good at generating turns out it is also very good at discriminating or predicting if a sequence is pathogenic or not."
>
> -- Eric Nguyen

This reveals a systems-level dynamic: if you only optimize for the obvious fix of blocking known bad sequences, you create a false sense of security. The system routes around your constraints by creating functional equivalents that look like benign, novel sequences to traditional scanners. The lasting advantage belongs to those who build models that understand the function of DNA, not just the spelling.

The 18-Month Payoff: Why Thinking in DNA Matters

Most AI teams focus on reading DNA. Nguyen’s team at Radical Numerics is pushing the frontier of reasoning via chain-of-thought in genomics. By training models to show their work--presenting progressively better RNA sequences with associated fitness scores--the model learns to navigate the trajectory of optimization on its own.

This approach requires significant upfront investment in alignment and post-training that most labs avoid. It is a classic example of immediate discomfort creating a lasting moat. While others struggle with context rot in long sequences, Nguyen’s team leveraged earlier innovations in long-context architectures to ensure their models maintain coherence across millions of base pairs.

> "If we show the model progressively better RNAs in a series of steps with its score... can the model continue that trajectory on its own? And then in the final step, does it self optimize to a point where it is like the best score it can get?"
>
> -- Eric Nguyen

This is not just a scientific curiosity; it is a shift in how biological discovery is performed. By moving from static sequence prediction to iterative, self-optimizing design, the system gains the ability to solve problems that were previously intractable, such as designing proteins for rare-earth mineral extraction or developing new classes of antimicrobials.

How the System Responds to Your Solution

A recurring theme in the conversation is the dual mandate. Nguyen emphasizes that because the defensive side has been lagging, the responsibility falls on those building the design capabilities to also build the safeguards. This creates a feedback loop: as design capabilities improve, the defensive models must also evolve.

The danger is not just intentional misuse by bad actors, but the unintentional risk of leakage, where experimental sequences meant for study escape containment. By embedding defense into the model architecture, the system becomes function-aware. Over time, this creates a defensive infrastructure capable of surveillance and attribution that can keep pace with the design frontier. The conventional wisdom--that we can keep the dangerous models in a box--fails when the underlying technology is essentially a mirror: the same engine that builds the cure can, if mismanaged, describe the pathogen.

Key Action Items

• Audit your sequence-matching assumptions: If your organization relies on simple blacklist-based filtering for biological sequences, recognize that this is vulnerable to functional obfuscation. (Immediate)
• Prioritize function-aware monitoring: Invest in or partner with labs developing models that evaluate the potential function of a sequence rather than just its similarity to known databases. (Over the next 6-12 months)
• Adopt the dual mandate framework: If building generative biological tools, integrate defensive and discriminative capabilities into the development lifecycle. Do not treat security as a post-hoc layer. (Immediate)
• Invest in domain-specific alignment: Move beyond base-model pre-training. Focus on mid-training and post-training alignment using Q&A-style structures to make genomic models actionable for specific clinical or diagnostic tasks. (Over the next 3-6 months)
• Embrace chain-of-thought for R&D: Experiment with training models on iterative, progressive sequences from low-fitness to high-fitness to see if they can recapitulate higher scores. This is a high-effort, high-reward strategy for biological design. (12-18 months)