This chapter addresses hardening AI-specific infrastructure components against model theft, data leakage, and cross-tenant contamination.
Untrusted AI models must be isolated in secure sandboxes, and sensitive AI workloads protected using trusted execution environments (TEEs) and confidential computing technologies.
| # | Description | Level |
|---|---|---|
| 4.1.1 | Verify that AI models execute in isolated sandboxes. | 1 |
| 4.1.2 | Verify that model artifact loading enforces an explicit allow-list of serialization formats that do not permit arbitrary code execution during deserialization. | 1 |
| 4.1.3 | Verify that workload attestation is performed before model loading to provide proof that the execution environment has not been tampered with. | 3 |
| 4.1.4 | Verify that confidential inference services protect model weights during runtime through isolated execution environments. | 3 |
AI-specific hardware components, including GPUs, TPUs, and specialized AI accelerators, must be secured.
| # | Description | Level |
|---|---|---|
| 4.2.1 | Verify that AI accelerator (GPU) firmware is version-pinned, signed, and attested at boot. | 2 |
| 4.2.2 | Verify that execution within a trusted execution environment (TEE) provides hardware-enforced isolation, memory encryption, and integrity protection. | 3 |
| 4.2.3 | Verify that AI accelerator (GPU) integrity is validated using hardware-based attestation mechanisms before each workload executes. | 3 |
| 4.2.4 | Verify that accelerator (GPU) memory is isolated between workloads through partitioning mechanisms with memory sanitization between jobs. | 3 |
| 4.2.5 | Verify that accelerator interconnects are restricted to approved topologies and authenticated endpoints. | 3 |
Distributed AI deployments, including edge computing, federated learning, and multi-site architectures, must be secured.
| # | Description | Level |
|---|---|---|
| 4.3.1 | Verify that edge AI devices authenticate to central infrastructure using strong authentication mechanisms. | 1 |
| 4.3.2 | Verify that models deployed to edge or mobile devices are cryptographically signed during packaging, and that the on-device runtime validates these signatures or checksums before loading or inference. | 2 |
| 4.3.3 | Verify that inference runtimes enforce process, memory, and file access isolation. | 3 |
| 4.3.4 | Verify that model weights and sensitive parameters stored locally are encrypted using hardware-backed key stores or secure enclaves. | 3 |
| 4.3.5 | Verify that models packaged within mobile, IoT, or embedded applications are encrypted at rest, and decrypted only inside a trusted runtime or secure enclave, preventing direct extraction from the app package or filesystem. | 3 |