Radical Numerics says DNA models are turning biosecurity into an AI arms race

Radical Numerics is building large AI models for biology while warning that the same systems that expand scientific capability could also widen the biosecurity threat. The company’s co-founder and chief executive, Eric Nguyen, made the case in an interview with Latent Space, arguing that defensive capabilities must keep pace with increasingly capable biological models.

Watch Desk analysis

What happened

Nguyen previously helped develop the Evo and Evo 2 genome models. Latent Space says related models were later used by a separate Arc and Stanford team to generate bacteriophage genomes that were synthesised into functional viruses. The interview presents that episode as evidence that genome-scale AI has moved beyond an intriguing research idea and into work with real-world biological consequences.

The discussion says Radical Numerics is now developing genome language models for a wider range of biological problems. Its models are described as working with DNA, RNA and proteins, with the company also pursuing other modalities such as three-dimensional protein structure and epigenetics.

Nguyen also describes an experiment involving RNA aptamers. The model was shown lower-performing examples in a sequence and asked to continue the pattern. According to his account, it reproduced some higher-scoring examples that had been held back from the training prompt. That is an attributed demonstration, not independent proof that the system can reliably design useful biological agents.

Why it matters

The practical point is not that AI has suddenly learnt biology in the way a person does. It is that long sequences, specialised training and multimodal systems may make biological design tasks easier to explore, automate and scale. A capability that helps researchers search for useful molecules can also create a harder problem for defenders trying to identify dangerous outputs or misuse.

That makes biosecurity an AI story for a very concrete reason: the technology is being applied to genetic and molecular information, not merely waving at an unspecified algorithm in the background. The stakes are unusually high, and the evidence deserves a little more than the industry’s usual confetti cannon.

Our read

Nguyen’s argument is persuasive as a warning about an emerging capability gap, but it remains an interviewee’s account of the technology and its trajectory. Readers should separate demonstrated research results from forecasts about what future systems might do. The sensible response is neither panic nor complacency. It is investment in testing, screening, secure access and independent evaluation before biological models become easier to deploy than they are to govern.

What to watch

• Independent replication:
  whether the reported aptamer result survives evaluation outside Radical Numerics.
• Model access:
  what safeguards, screening and monitoring accompany increasingly capable genome models.
• Defensive capacity:
  whether public-interest researchers can test systems without lowering barriers to misuse.
• Company progress:
  whether Radical Numerics publishes concrete models, benchmarks and safety procedures rather than broad ambitions. Latent Space’s interview gives the debate a sharper shape: AI is not only accelerating biological discovery, it may be changing the balance between offence and defence. That is a race worth watching, even if nobody should be handing out track medals yet.

Discussion spark: Should advanced biological models be developed openly so defenders can keep pace, or should access be restricted until independent safety testing catches up?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.