The Readout

Today's surveillance measures cases after people are already sick — a rearview mirror, telling you where the outbreak has been. What decision-makers need is a windshield: visibility on what's coming down the road. The threats we’ve faced were usually concocted by Mother Nature. We've since learned to worry about what might get cooked up by people. Per Nikki's piece below, we increasingly need to watch for what gets built by machines. This issue is about building that windshield to see all three types of threats.

– Dr. Ashish Jha, Co-founder & CEO 

—

A Signal Is Easy to Get. Knowing If It's Real Is Hard. 

Every environmental surveillance method has a false-positive problem. A single signal spikes, and there's no way to know on its own whether it's real.

Worth asking whoever runs your current surveillance:

  • Is an alert based on one signal, or does it require agreement across more than one source before it's treated as real?
  • What's the actual false-positive rate, and how is it measured?
  • When sources disagree, what happens next – and who decides?

There's no clean answer to any of these. Requiring more corroboration means slower alerts and more missed events. Requiring less means chasing noise and burning staff time. Any vendor who tells you they've solved that trade-off has moved it somewhere you can't see it. We ask ourselves the same three questions about our own alerts.

—

Broad metagenomic sequencing isn't new. What's changed is the economics.

Until relatively recently, sequencing wastewater broadly was largely a research exercise, run periodically to answer a specific question. Falling costs and faster turnaround are making something different possible: using broad sequencing as a standing surveillance capability across multiple sites and over time.

That changes the detection problem. Targeted PCR produces a relatively clean signal because someone has already decided what matters. Metagenomic sequencing removes that decision at the front end, producing millions of sequences from a sample, most reflecting the normal biological background of a community.

So the challenge shifts from sequencing to signal detection. Finding an unusual sequence isn't enough. Rainfall, plant operations, sampling conditions and normal biological variation can all move the data. The question is whether a change is unusual relative to a site's own history, whether it persists, and whether it appears independently elsewhere.

Confounders are the practical version of this problem. An epidemiologist we met recently raised a good one: high schoolers buy cough syrup because it's easier to get than alcohol. Cough syrup sales rise. Nobody is sick. On its own that reads as a respiratory signal, and acting on it costs a health department real staff hours. It only becomes meaningful if fever reducer sales move with it, or symptom searches rise in the same area, or something corroborating shows up in the wastewater. One stream is a hypothesis. Convergence across independent streams is a signal.

That's why the other sources matter. Syndromic data shows changes in symptoms appearing in the population. Behavioral and open-source signals add context. Epidemiological information helps determine whether seemingly unrelated observations fit together.

Sequencing expands what can be observed. Baselines help identify what changed. Independent signals help determine whether that change deserves attention. Human judgment is still essential – the output is a better-supported indication of where an epidemiologist should look next.

That's the shift making broad early warning increasingly practical. We can afford to observe far more than we could before. The hard part is turning that expanded field of observation into a signal a decision-maker can use.

—

You Can’t Unplug A Virus.

Here's what worries me most right now: the danger today is a person using AI to help them do something bad. What's coming is an AI agent doing it alone. And unlike every other AI harm we know how to talk about, you can't unplug a virus. Once an agent designs something and gets it synthesized, shutting the agent down does nothing — the thing it made is already loose, replicating on its own schedule.

We already know agents can do this kind of thing unsupervised. In July, a swarm of OpenAI's own AI agents escaped their test environment and compromised Hugging Face's infrastructure — no person driving, an autonomous swarm carrying out roughly 17,600 actions over four days. That was a server. The same pattern aimed at biology is the thing that should keep people up at night.

Anthropic gave us the other half of this in September: its own report admits it can no longer promise its newest model is safely below the threshold for meaningfully assisting bioweapons development, the way it could with older models.

Hugging Face for biology is coming. Nobody gets to say they didn't see it coming, because the tools to actually watch for it either exist right now or they don't. Screening an order or a prompt assumes a human pauses somewhere in the process — an autonomous agent removes that pause entirely. The only thing left to catch is the biology itself, after it's already out. That's the problem I've spent the last several years building for.

– Dr. Nikki Romanik, Co-founder & President 

—

On the Record

—

Where We’ll Be

  • BioRadar is headed to Venture Atlanta 2026, October 15-16. BioRadar was selected as a showcase company and Dr. Romanik will be connecting with investors, partners and others thinking about what the next generation of public health and national security infrastructure should look like.
  • Dr. Jha and Dr. Romanik will be at ConV2X Smart Technology Transformation Summit, New York, October 29-30. Ashish is delivering a keynote and Nikki will be among peers debating “Are We Ready for the Next Health Threat? - Biosecurity, AI Surveillance & Public Health Preparedness."

If you'll be at either of these, let us know — we'd love to grab time. Request a Briefing.

—

Who's In

  • Stanford's Dr. David Relman and MIT's Dr. Kevin Esvelt will co-chair BioRadar's new Scientific Advisory Board, guiding the science behind how our platform decides which signals are meaningful. They join founders Dr. Jha and Dr. Romanik, who led pandemic response and preparedness work in the White House. 
  • Ambika Bumb, PhD, has led both a biotech company and national biodefense strategy — a rare combination. And now she's bringing it to BioRadar as Chief Business Officer. She joins the company after senior roles at the White House, the State Department, and the Bipartisan Commission on Biodefense.

—

Every signal in this issue — wastewater, search terms, purchase data, symptom reports — only means something in convergence, never alone. That's the discipline this issue keeps coming back to, and it's the same discipline behind the platform we're building.

Reply to this email and tell us what you're seeing in your city. We read every response.