BioAI Safety & Security (BioAISS)

Biosecurity is a core element of national security, and rapidly advancing AI has lowered the barrier for adversaries to design biological and neurological agents. AI-designed threats need not resemble anything in our reference databases, which defeats signature- and homology-based detection. Reasoning over the physical interaction surface rather than sequence identity gives us a detection and triage pathway that does not depend on prior observation of the agent, and the same model accelerates the countermeasure side: target identification, antibody developability, and polyspecificity risk assessment at the speed a response timeline actually requires.

Our group developed SurfGraphPro, a foundation model that learns protein surface geometry and physicochemical composition directly. SurfGraphPro exceeds state-of-the-art accuracy by 8% with an 18-28x speedup, generating surface-informed embeddings and predicting protein-protein interaction sites in seconds to under a minute per protein. This moves surface-based target screening from a per-protein exercise to a genome-scale capability. SurfGraphPro complements our protein language model work on pathogen classification and biosurveillance.

We are leveraging SurfGraphPro to perform systematic, proteome-scale characterization of protein interaction surface, which accelerate both detection of novel biothreats and designing medical countermeasures against them.

The biological capabilities of AI are at once among the most promising for helping humanity, the most potentially terrifying if misused, and the most challenging to accurately measure. SecureBio builds tools to make these measurements, with the aim of reducing the probability of catastrophic misuse.

Here, we present an overview of our work to date, exploring benchmarks that demonstrate how frontier models can run liquid-handling robots, inexpensively reproduce published biological AI models using just an abstract, and troubleshoot virology better than most expert virologists.

We go on to look in depth at our recent explorations of how biosafety mitigations might function in an open-weights world, where there are no system-level safeguards: how do interventions like pre-training data filtering and post-training unlearning hold up to targeted attack?

Finally, we provide a preview of our upcoming work to extend critical biosafety-relevant assessments from the digital to the molecular.

The biological capabilities of AI are at once among the most promising for helping humanity, the most potentially terrifying if misused, and the most challenging to accurately measure. SecureBio builds tools to make these measurements, with the aim of reducing the probability of catastrophic misuse.

Here, we present an overview of our work to date, exploring benchmarks that demonstrate how frontier models can run liquid-handling robots, inexpensively reproduce published biological AI models using just an abstract, and troubleshoot virology better than most expert virologists.

We go on to look in depth at our recent explorations of how biosafety mitigations might function in an open-weights world, where there are no system-level safeguards: how do interventions like pre-training data filtering and post-training unlearning hold up to targeted attack?

Finally, we provide a preview of our upcoming work to extend critical biosafety-relevant assessments from the digital to the molecular.