Summary: Detection is only worthwhile if it’s fast enough to trigger an effective response. The OpenAI Foundation has granted SecureBio Detection $17.2M to reduce our end-to-end time from fourteen to three days, expand our collection footprint, and further validate our detection system.
The detection gap
We remain concerned about the world’s vulnerability to pathogens that could spread widely before detection. Traditional list-based detection allows you to track known threats, but a pathogen could be engineered to not match anything on those lists. We view metagenomic sequencing as the best opportunity for pathogen-agnostic detection, going beyond lists of known infectious agents to capture the wide range of nucleic acids in a sample.
With support from Coefficient Giving and others we collaborated with our CASPER partners to massively scale up wastewater sequencing, expanded to nasal swab sequencing, built and validated pipelines for detecting engineered sequences, and began running an AI-driven flag analysis system in production. We’ve shown that pathogen-agnostic detection works, but we still need to show that it can be done fast enough to provide meaningful protection.
What we’re planning to do
The OpenAI Foundation is funding a focused push to demonstrate that speed is possible here. Over the next year we will:
Cut the time from collecting a sample to actionable results down to three days.
Increase our average weekly throughput to 80B read pairs.
Expand our nasal swab program to three additional cities.
Most of the speed improvements come from integration and operations work rather than new research. Major projects here will be optimizing our sample transport logistics, automating and parallelizing library prep, and restructuring our data pipelines to take better advantage of burst compute. The throughput increase is different: our lab has already built the capacity to process more than 80 samples per week, but we’ve been running well below that because of the expense of flow cells and other consumables.
We will also continue broadening our defenses by developing and validating detection methods. On the lab side, we’ll run end-to-end sensitivity validation to ensure that no classes of virus are systematically missed. On the computational side, we’ll red-team our system, both in-house and contracting with third-party consultants. We’ll identify and close gaps an attacker might try to exploit.
Why this matters now
The risk landscape is changing quickly. In SecureBio’s own evaluations, frontier models now score higher than every human expert we’ve tested in their area of expertise. Cybersecurity risk is farther along this track and shows how escalating capabilities can change what real-world threats look like. A year ago, AI was lowering barriers to cybercrime but still with humans directing each step; a few months later, Anthropic reported: “We believe this the first documented case of a large-scale cyberattack executed without substantial human intervention.” And then, in just the last month, three different AI firms (OpenAI, Anthropic, Meta) disclosed that their models are now sufficiently capable of sustained autonomous effort that they compromised real systems belonging to other companies during cybersecurity evaluations. Measured capabilities on cyber and biology tasks grow together, so we could well see similar real-world capability growth in bio, and it might come quickly.
The world will soon need robust defenses in place that don’t depend on any single company’s safeguards, and pathogen-agnostic early warning is a critical component.
How this interacts with SecureBio’s model evaluation work
SecureBio also evaluates biosecurity risks from frontier AI models. We were clear that this grant would not influence what we say about OpenAI’s models, or the models from any other firm, and the OpenAI Foundation is fully aligned with that principle. Specifically:
SecureBio’s model evaluations, including pre-release assessments for biosecurity risk, are under our AI team. This team operates independently from Detection, with separate leadership, budgets, and deliverables.
This grant is restricted to where it may only fund our Detection work, and it places no constraints on what any part of SecureBio can research, publish, or say.
Our AI team’s policy is to evaluate models solely on their merits.
What comes next
This grant will let us bring our system much closer to what’s needed. A system at the scale that would robustly protect a country, however, will require significantly more money and effort. What this grant lets us do is show that fast pathogen-agnostic detection is achievable, give philanthropic and government funders a clearer picture of what a scaled system would cost and deliver, and continue building towards larger-scale systems that would provide this robust protection.
We’re grateful to the OpenAI Foundation for the funding, and to our collaborators across the CASPER network. If you’re working on something adjacent and would like to talk, get in touch.
