Data in use is information that is actively being processed by a system, such as when a program decrypts a file, reads a prompt into memory, or runs a model inference. It is neither sitting on disk nor traveling across a network; it is in memory, registers, caches, or accelerator buffers while computation happens.
This state matters because encryption at rest and in transit does not fully protect plaintext during processing. Attackers may target that moment with memory scraping, malicious insiders, hypervisor abuse, side-channel attacks, or compromised CPUs and GPUs. Defenders reduce this risk with confidential computing, secure enclaves, hardware-backed isolation, attestation, and strict secret handling so sensitive inputs, keys, weights, and training data stay protected while they are used. In AI systems, protecting data in use is critical because prompts and model assets are often most exposed during runtime.