Definition
A hardware-based protection that keeps data and AI model weights encrypted even while they are actively being used in a computer's memory, not just when stored or transmitted. It relies on a verification step called 'attestation' that checks the hardware hasn't been tampered with before releasing sensitive data into it — so in theory, even the cloud provider running the machine can't see inside.
Why it matters
Companies running sensitive AI workloads on rented cloud hardware depend on this guarantee to protect proprietary models and data from the infrastructure operator itself; if the verification step can be tricked, the entire protection collapses without anyone noticing.