Trusted Execution Environment (TEE)

Protects data in use by ensuring it cannot be read by unauthorized processes, even if the host OS or hypervisor is compromised. Inside the TEE, data is decrypted only within the CPU package.

• Mechanism: Hardware-managed memory encryption engines transparently encrypt and decrypt cache lines.
• Real-world example: An AI model loaded into a TEE cannot have its weights extracted by a malicious system administrator.
• Key distinction: This is distinct from data-at-rest encryption (storage) and data-in-transit encryption (TLS).

Guarantees that code and data inside the TEE cannot be modified by external software. Any unauthorized write attempt is blocked at the hardware level.

• Mechanism: Memory pages assigned to the TEE are integrity-protected via cryptographic hashes stored in a reserved area of DRAM.
• Real-world example: An audit log signing key stored in a TEE cannot be tampered with by a rootkit.
• Attack mitigated: Prevents active physical attacks like DRAM row-hammer from corrupting secure computations.

Provides cryptographic proof of the exact software stack loaded inside the TEE. A remote party can verify the TEE's identity and that it is running unmodified, trusted code before sending secrets.

• Mechanism: The CPU generates a signed report (quote) containing a hash of the TEE's initial memory state.
• Real-world example: A data provider sends sensitive financial data to a TEE only after verifying its attestation report against a known good hash.
• Key standard: Follows the IETF Remote ATtestation ProcedureS (RATS) architecture.

Creates a strict physical boundary between the secure world and the normal world. The CPU enforces this separation, preventing any software from bridging the two domains.

• Mechanism: A hardware bit (e.g., ARM TrustZone's NS bit) or a dedicated CPU mode (e.g., Intel SGX's enclave mode) tags transactions on the system bus.
• Real-world example: Even a kernel-level debugger cannot set a breakpoint inside a TEE's protected memory region.
• Scope: Isolates not just the CPU, but also interrupts, caches, and direct memory access (DMA) paths.

Allows a TEE to encrypt data and bind it to a specific device and software identity, ensuring data can only be decrypted by the exact same TEE application in the future.

• Mechanism: Encryption keys are derived from a CPU-specific fuse key and the TEE's software measurement.
• Real-world example: An AI audit log is sealed to a specific TEE, preventing an attacker from copying the encrypted log file to another machine for offline decryption.
• Benefit: Provides secure persistence without requiring the user to manage complex key hierarchies.

Reduces the attack surface by excluding the host OS, hypervisor, and device drivers from the security perimeter. Only the TEE's verified code must be trusted.

• Mechanism: The CPU's microcode and the TEE application itself are the only components in the TCB.
• Real-world example: A vulnerability in the Linux kernel does not compromise the integrity of a confidential computing enclave.
• Design principle: This radical reduction in complexity makes formal verification of the security model practically achievable.