Whitebox AI Labs

Insurers, law firms and defence buyers are being asked to underwrite AI systems whose runtime behaviour nobody can audit. Whitebox scores a system against a published standard and issues a certificate with a deployment envelope, so the risk can be priced before procurement rather than discovered afterwards.

Frontier models have been observed resisting shutdown and behaving autonomously under test, across every major provider rather than at one lab. That is a governance problem long before it is a research one: an insurer asked to cover an AI-driven process, or a law firm asked to sign off on one, has no standard way to measure what the system will do at runtime, and no document to point at when asked how they knew.

Whitebox builds the measurement and the paperwork around it. Its first product scores any AI system against a published standard and issues a certificate with an explicit deployment envelope, aimed at the regulated buyers who need it rather than at the labs building the models. The second product moves from measuring the risk to containing it at runtime. The patent family behind both is the reason a fast follower cannot simply copy the approach.