AI protein design tools can generate sequences that read as harmless to homology-based biosecurity screens, while folding into dangerous structures. Foldguard screens at the structural level, deployable inside AI labs or at the point of synthesis, where that gap closes.
An AI-redesigned toxin can carry almost no resemblance to any known sequence while folding into the same dangerous shape. Sequence screening compares it in the space where it hides. Foldguard compares it in the space where it's exposed.
Illustrative schematic. Similarity shown qualitatively.
The same structural comparison engine deploys at either end of the design-to-synthesis pipeline, wherever the order flow actually passes through.
Deployed in-VPC alongside protein design and biology-capable models, Foldguard acts as a structural safeguard on model outputs, checking generated sequences against known hazardous folds before they're ever returned, inside the lab's own environment.
Deployed in the synthesis pipeline, Foldguard adds structural screening on top of existing sequence checks, helping providers stay ahead of tightening screening requirements and flag the threat class that homology alone misses.
Each sequence, whether freshly generated or submitted for synthesis, is run through structure prediction to generate the shape it would produce once expressed.
That predicted fold is compared against a database of known hazardous structures using structural alignment, not sequence identity.
Matches above threshold are flagged for review before the sequence is returned or synthesized, with the matched structure attached for context.
Two things are happening at once: biosecurity regulation is moving from best practice toward baseline, and frontier labs are increasingly recognizing the importance of building safeguards into how their models are released. Foldguard is built to serve both.
Federal frameworks now make synthesis screening a condition of life-sciences funding, the screening window is dropping toward shorter sequences, the definition of sequences of concern is expanding, and bipartisan legislation introduced in 2026 points toward mandatory screening. Structural methods are how providers stay ahead of where the rules are heading.
Frontier labs are increasingly recognizing the importance of biosecurity safeguards for biology-capable models, and need concrete tools to put them in place. Foldguard gives labs a structural screen on model outputs they can run in their own environment, turning that recognition into a working control at the point of generation.
Regulators, red-teams, and published research have all converged on the same blind spot: sequence screening alone can miss AI-designed threats. A structural layer is an auditable, demonstrable response to the exact risk class now under scrutiny.
While the issue is not new, the pace of progress in artificial intelligence is.
One commercial biosecurity screening tool flagged less than a quarter of the AI-redesigned toxin sequences in a 2025 red-team study run with four major DNA synthesis providers.
In 2025, Microsoft researchers led by Bruce Wittmann, working with IBBIS and four commercial DNA synthesis providers, generated tens of thousands of AI-redesigned variants of known toxins and tested them against production screening software. Several tools missed the majority of the redesigned sequences.
The screening industry has since patched against that specific study's outputs. But the underlying gap, that sequence comparison can't see a fold it doesn't recognize, is structural to the approach itself, not a bug in one version of the software. Foldguard is built to close that gap wherever a sequence first takes shape, whether that's inside a model or at the point of synthesis.
Foldguard is an early two-person company. One side builds the science; the other builds the company and the partnerships to get it used.
Computational biologist with a PhD from the University of Arizona, with research experience including work for the U.S. Department of Energy. Leads the science behind Foldguard, from structure prediction to structural comparison against known threats.
Leads partnerships, fundraising, and strategy. Background in product and data operations.
We're early, and talking with frontier AI labs, synthesis providers, and biosecurity funders to shape structural screening around how it'll actually be used.
Get in touch