Radical Numerics launches offline biodefense AI that scores pathogen risk on a laptop

Radical Numerics on Wednesday unveiled Omnii Defense, an offline AI biodefense model designed to flag both natural and AI-engineered pathogens. The company says the system can run on a laptop connected to a USB gene sequencer, a setup that could matter for field screening, public-health labs and other settings where internet access is limited or unwanted.

The launch adds a new entrant to a fast-moving corner of AI biology: tools meant not to design pathogens, but to spot risky sequences before they spread or are acted on. Radical is limiting access to vetted external testers for now, and it says it is working with national labs, government agencies and other evaluators.

What Omnii Defense is supposed to do

Omnii Defense is built to analyze genetic sequences and return multiple signals at once. According to Radical, it predicts pathogenicity, estimates the likelihood that a sequence was AI-generated, and identifies the closest taxonomic group. It then produces a threat score from 0 to 1 that combines what the company describes as unnaturalness and functional predictions.

That structure matters because older biodefense tools, Radical says, mostly work by matching sequences against libraries of known organisms. Omnii is intended to look for patterns instead, including cases where the genetic spelling has changed enough to evade simple lookup-based screening.

Why does an offline model matter?

Because the company is pitching Omnii Defense for situations where connectivity is a weakness, not a feature. A system that runs locally on a laptop and a USB gene sequencer can, in theory, be used at the edge of a public-health response, in a lab with limited infrastructure, or in a security-sensitive environment where keeping data off the internet is important.

That offline design also signals how Radical wants the product to be read: not as a cloud service for casual users, but as a controlled tool for vetted testers and institutional evaluators. The company has not opened the model to broad public access.

Why it matters

The practical question is whether AI can improve biodefense faster than it lowers the barrier to bioengineering. That is the tension surrounding Omnii Defense and similar systems: the same pattern-recognition methods that may help defenders spot manipulated sequences could also accelerate dangerous experimentation elsewhere.

For readers outside the biosecurity field, the stakes are easier to frame in operational terms. Public-health teams and national labs are under pressure to detect threats earlier, but current screening methods are often built around known references. Radical’s pitch is that an AI model trained to read sequence-function relationships can add another layer of scrutiny when the reference itself is missing or altered.

For the market, the announcement is also a signal that AI biology companies are moving beyond demonstration projects and into applied defense use cases. The company’s involvement with national labs and government agencies suggests a path toward institutional evaluation rather than consumer adoption, and possibly toward future regulatory scrutiny if the tool is used more widely.

How did Radical test it?

One of the clearest examples in the fact sheet involves Bundibugyo virus. Radical says the model still flagged some sequences even after the species was removed from the training context, and again after the genus was removed. In that experiment, non-AI tools produced essentially no risk signal.

That result is not the same as proving a real-world threat. But it does suggest that the model can continue to surface warnings when straightforward library matching would fail. In a field where bad actors may alter genetic sequences precisely to avoid known signatures, that kind of resilience is central to the company’s pitch.

Still, the test result should be treated carefully. A flagged sequence is a signal, not a verdict. And any screening system that claims to identify engineered biology needs to be checked against real data, real workflows and edge cases that can produce false positives or false negatives.

What the experts said

Radical’s leadership frames Omnii Defense as a response to a widening gap between offensive capability and defensive readiness. Co-founder and CEO Eric Nguyen described biosecurity as lagging behind other AI-adjacent defense areas, telling WIRED:

“We felt that the biosecurity space was far behind on the defense side.”

Christine Parthemore, Radical’s head of biodefense, went further in characterizing the product, calling Omnii a

“step change”

. Her wording suggests the company sees the model not as an incremental update but as a different class of screening tool.

Outside the company, the reactions in the fact sheet are more guarded. Bronwyn MacInnis, director of pathogen surveillance at the Broad Institute of MIT and Harvard, said:

“We absolutely need to be vigilant about what AI is enabling”

She also warned that detection is not the same as proving real consequence or engineered intent, a distinction that matters in any operational biosecurity setting.

Eric Nguyen’s comparison is the bluntest one. He framed biodefense as an AI arms race similar to cybersecurity, saying the field has been far behind on the defense side. That is a familiar claim in emerging-tech circles, but in this case it carries a specific policy meaning: if offensive tools advance faster than screening, public institutions may have less time to react to manipulated pathogens.

Background: how Radical Numerics got here

Radical Numerics was founded in 2025 in San Francisco, and its founders previously worked on Evo and Evo-2 at Stanford. Those projects sit in the broader category of biology foundation models, systems designed to learn from biological sequence data and infer patterns that may not be obvious to conventional software.

That background helps explain why Radical is now talking about biodefense in AI terms. The company’s approach appears to build on the idea that sequence models can learn relationships between genetic code and biological function, not just memorize known examples. In theory, that gives them a chance to catch novel or modified threats that older tools might miss.

The broader field has already wrestled with the dual-use problem. Frontier AI labs have warned that AI could be used to help design dangerous bioweapons. At the same time, defenders have argued that AI can also help close gaps in surveillance, screening and triage. Omnii Defense lands squarely in that debate.

There is also a practical reason the field is paying attention: bacteriophages are viruses that infect bacteria, and they can be engineered to target specific strains. In October, Radical said Omnii helped create custom bacteriophages for drug-resistant E. coli, a claim that suggests the same platform may be able to support both therapeutic and security-sensitive workflows. That work involved a sick patient and special FDA permission, according to the fact sheet.

How should readers interpret the phage work?

As an indicator of capability, not a blanket safety certificate. The bacteriophage example shows Radical positioning Omnii as a useful biological design assistant, but it does not settle the question of how well the same system handles defensive screening, nor does it answer how safely the model could be deployed more broadly.

It also underscores why governance matters. A model that can contribute to customized biological tools in one setting may require different controls when it is used to flag pathogen risk in another.

The Bundibugyo experiment and what it suggests

The Bundibugyo virus test is the closest thing in the fact sheet to a stress test. Radical says the model still identified some sequences after the species-level context was removed and even after the genus-level context was removed. The company also says non-AI tools gave essentially no risk signal in that same experiment.

In plain terms, the comparison is meant to show that Omnii Defense can retain suspicion even when obvious taxonomy cues are stripped away. That is the kind of behavior a defender wants if an adversary tries to disguise a sequence by changing enough of its genetic “spelling” to dodge a database match.

But the test also hints at a future validation challenge. If a model can flag modified or unfamiliar sequences, how often does it do so accurately? How often does it misread harmless samples? And can it tell the difference between a genuinely engineered threat and a benign sequence that just happens to look odd to the model?

Those are not academic questions. In a clinical or public-health setting, false alarms can waste time and resources, while missed threats can have far larger consequences. The value of any AI biodefense tool will depend on how it performs under conditions that are messier than a controlled benchmark.

Who is behind the company?

Radical’s leadership comes from a mix of AI biology and biodefense experience. Nguyen and the company’s founders previously worked on Evo and Evo-2 at Stanford, placing them in the lineage of researchers who helped popularize the idea that foundation models can be useful in biology.

Christine Parthemore is the company’s head of biodefense, a role that appears to bridge technical development and security strategy. Her use of the phrase “step change” signals that Radical wants to be seen as moving beyond standard pathogen databases.

There has also been outside scrutiny. Harvard geneticist George Church resigned as a Radical adviser last week after a disclosure dispute over a Tabula Bioscience partnership, according to the fact sheet. That episode does not directly bear on Omnii Defense’s technical performance, but it does show that the company’s scientific relationships are already under attention.

What happens next

For now, the key unknowns are less about the announcement itself and more about what follows from it. Radical says access is limited to vetted external testers, so the next phase likely depends on how those evaluators respond and whether the company can show reliable results outside its own demonstrations.

Here are the main things to watch:

• Independent validation of Omnii Defense’s performance on real or representative pathogen data.
• Whether national labs or government agencies publicly endorse operational use.
• How broadly Radical expands access beyond vetted testers.
• Whether the model can consistently distinguish genuine engineered risk from harmless anomalies.
• How policymakers and biosafety reviewers respond to AI systems that can both aid and complicate biosecurity.

Another open question is whether tools like Omnii Defense will remain mainly defensive screening systems or become part of more routine lab infrastructure. If they do, the consequences could extend beyond specialized biodefense work into areas like diagnostics, surveillance and regulated research.

That is why the announcement matters beyond one company. Radical is not simply marketing a new model; it is offering a view of what biosecurity might look like if AI becomes part of the front line. The promise is faster detection and better pattern recognition. The risk is that the same technology may make the underlying science easier to manipulate.

For now, Omnii Defense is still in a controlled phase, with evaluation rather than broad deployment ahead. But the direction is clear enough: AI is moving deeper into the infrastructure of biodefense, and the debate around how to use it safely is only getting sharper.