Trusted Execution Environment (TEE)

A core design principle of TEEs is the radical reduction of the Trusted Computing Base (TCB)—the set of all components that must be trusted for security to hold. In a traditional stack, the TCB includes the hypervisor, host OS, and firmware. A TEE shrinks the TCB to the enclave application, the processor package, and a thin security monitor.

• Eliminates the cloud provider's hypervisor from the trust boundary.
• Simplifies security auditing and formal verification efforts.
• Reduces the attack surface for zero-day vulnerabilities in the infrastructure layer.

Enclave sealing allows a TEE to encrypt data for persistent storage in a way that binds the ciphertext to a specific enclave's identity. The sealed data can only be decrypted by the exact same enclave code running on the exact same CPU. This protects sensitive state at rest, such as model parameters or user data, from the host file system.

• Binds persisted AI model checkpoints to a specific, attested inference service.
• Prevents an attacker with root filesystem access from reading sealed secrets.
• Enables secure stateful enclaves that survive process restarts.

Modern TEEs incorporate hardware and microcode defenses against side-channel attacks that attempt to infer secrets by observing timing variations, power consumption, or cache access patterns. Techniques include cache partitioning, execution path hardening, and speculative execution controls to mitigate vulnerabilities like Spectre and Meltdown within the enclave boundary.

• Protects cryptographic keys from extraction via cache-timing analysis.
• Ensures AI model architecture cannot be reverse-engineered through memory access patterns.
• Continuously updated via microcode patches to address newly discovered attack vectors.

The process of cryptographically verifying the identity and integrity of a Trusted Execution Environment, ensuring the hardware, firmware, and software have not been tampered with.

• Produces an enclave measurement—a cryptographic hash of the initial code and configuration
• Relies on a hardware root of trust to sign attestation reports
• Enables remote parties to verify they are communicating with a genuine, untampered enclave
• Critical for confidential AI: verifies model provenance before releasing decryption keys

An extension of AMD Secure Encrypted Virtualization that adds memory integrity protection and prevents malicious hypervisor-based attacks like data replay and memory remapping.

• Encrypts entire virtual machine memory with a key unique to that VM
• Secure Nested Paging prevents hypervisor from remapping guest memory
• Provides attestation report signed by the AMD Platform Security Processor
• Enables lift-and-shift of existing applications into confidential VMs without code changes

A hardware-based security capability that protects data in use within NVIDIA GPUs, creating an isolated execution environment for AI workloads to prevent unauthorized access during computation.

• Extends TEE protection to GPU-accelerated AI training and inference
• Encrypts data in transit across PCIe bus between CPU and GPU
• Supports secure GPU attestation to verify firmware integrity before offloading sensitive computation
• Enables multi-party AI collaboration where model weights and training data remain confidential

A virtual machine instance backed by hardware-based memory encryption, ensuring that data remains encrypted while in use and is isolated from the cloud provider's hypervisor.

• Supported by Intel TDX, AMD SEV-SNP, and ARM CCA technologies
• Enables existing applications to run confidentially without modification
• Protects against privileged insider threats at the cloud provider level
• Foundation for confidential inference services and sovereign cloud deployments

A mechanism that allows a Trusted Execution Environment to encrypt data for persistent storage, binding it to a specific enclave identity so it can only be decrypted by the same application on the same platform.

• Derives encryption keys from the enclave's unique cryptographic identity
• Prevents sealed data from being read by other enclaves or the host OS
• Supports sealing to enclave identity (same enclave on same platform) or sealing to signing identity (same enclave author across platforms)
• Essential for confidential persistent storage in stateful AI workloads