For frontier labs and the AI safety community: an independent channel that carries technical findings to the institutions deciding how American AI gets fielded and governed.
Frontier labs are producing capabilities with direct national security implications faster than the government can absorb them. Competitive and legal constraints limit what the labs can share, both with the government and with one another, and there is no neutral channel between the people building these systems and the people responsible for fielding and governing them.
Frontier Security Institute is that channel: an independent party with no commercial stake, able to assess frontier capabilities, translate them into terms the government can use, and bring both sides to the same table.
Independent analysis, editorial standards on both sides, and attribution on our partners’ terms.
Non-commercial assessment of frontier model capabilities in the scenarios that matter to national security, produced by an evaluator with no equity in the outcome, so the findings are trusted on both sides of the table.
Turning technical findings into the forms operators, program offices, and acquisition leaders can act on, without losing technical fidelity.
A setting where the labs and the National Security Enterprise can speak candidly and off the record, and where the labs can compare findings they otherwise cannot share.
Working with the labs on the technical, legal, and operational means to protect models, weights, and the people who build them, and connecting that work to the existing authorities available.
Our Model Security program is built around this work with the labs.
See our programs →Frontier Security Institute focuses on the National Security Enterprise. Our work extends the Center for AI Safety’s mission of reducing societal-scale risks from artificial intelligence.
Center for AI Safety →Evaluations, red-teaming, joint convenings, and technical advisory: if your work shapes what frontier AI can do, ours helps determine how it gets used.