The Breach That Changes the Model

We have spent a decade learning to defend the inputs to artificial intelligence and almost no time defending the model itself. Weight tampering is the breach you will not detect, because the system keeps answering and the answers look fine. The only durable defence is a record of what the model was and what it did, signed before it acts and verifiable without trusting the vendor.