Skip to content

Confidential Computing & TEEs

Confidential Computing is the protection of data in use by performing computation in a hardware-based, attested Trusted Execution Environment (TEE).

What is a TEE?

A Trusted Execution Environment (TEE), or Enclave, is a secure area of a main processor. It guarantees that code and data loaded inside are protected with respect to confidentiality and integrity.

Key properties include: - Data-in-use encryption: Memory used by the enclave is encrypted. - Isolation: The OS, Hypervisor, and other processes cannot peek inside the TEE. - Attestation: The hardware provides a cryptographic proof that a specific piece of code is running on genuine hardware.

Why TEEs Matter for AI

Traditional AI deployments have a trust problem:

Scenario Without TEE With TEE
Data privacy You must trust the operator won't log your data Hardware guarantees data isolation
Model integrity You must trust the model hasn't been modified Cryptographic hash proves exact code running
Guardrail enforcement You must trust safety checks are enabled Attestation proves auditors are active
Geographic compliance You must trust data stays in-region Hardware proves execution location

With TEEs, you don't need to trust — you can verify.

Supported Hardware

Lucid supports the following TEE technologies:

Technology Vendor Cloud Availability
Intel TDX Intel Azure, GCP
AMD SEV-SNP AMD Azure, AWS, GCP
AWS Nitro Enclaves AWS AWS
Intel SGX Intel Azure (legacy)

All provide similar security guarantees but differ in implementation details. Lucid abstracts these differences, allowing you to deploy to any supported platform.

Attestation in Lucid

Lucid leverages remote attestation to provide users with an AI Passport. This passport contains:

  1. Hardware Quote: Proof it's a real TEE (e.g., Intel TDX, AMD SEV-SNP, AWS Nitro).
  2. Software Measurement (MRENCLAVE): A hash of the exact code running.
  3. Public Key: Used to verify subsequent results signed by the enclave.

Attestation Flow

sequenceDiagram
    participant User
    participant Lucid Verifier
    participant TEE Enclave
    participant Hardware Root of Trust

    User->>Lucid Verifier: Request AI Passport
    Lucid Verifier->>TEE Enclave: Request attestation report
    TEE Enclave->>Hardware Root of Trust: Generate signed quote
    Hardware Root of Trust-->>TEE Enclave: Signed attestation
    TEE Enclave-->>Lucid Verifier: Attestation report + measurements
    Lucid Verifier->>Hardware Root of Trust: Verify signature chain
    Hardware Root of Trust-->>Lucid Verifier: Verification result
    Lucid Verifier-->>User: AI Passport (if valid)

What Gets Attested?

Component Measurement What It Proves
Model Hash of model weights Model hasn't been tampered with
Auditors Hash of auditor code Exact safety checks running
Runtime Hash of execution environment No malicious code injection
Configuration Hash of policy config Correct settings applied

Verifying Attestation

You can verify TEE attestation using the CLI:

# Verify a specific environment
lucid verify environment env-abc123

# Check attestation details
lucid passport show pass-001 --attestation

Or programmatically:

import requests

# Fetch AI Passport
response = requests.get("https://verifier.lucid.sh/v1/passports/pass-001")
passport = response.json()

# Check attestation status
if passport["hardware_attested"]:
    print(f"TEE Type: {passport['tee_type']}")
    print(f"Measurement: {passport['mrenclave']}")
else:
    print("Warning: Not hardware attested")

Serverless Mode

In serverless mode, Lucid manages shared TEE resource pools. The same hardware-backed security guarantees apply — TEE memory encryption and attestation verification ensure equivalent isolation to dedicated deployments. See the Deployment Modes guide for details on getting started with serverless.

Mock Mode

Since TEE hardware is specialized and not available on most development machines, Lucid provides a Mock Mode for local development. It simulates the attestation handshake using software-based keys, allowing you to build and test your logic before deploying to cloud infrastructure.

Enabling Mock Mode

Mock mode is enabled via the TEE_PROVIDER=MOCK environment variable on the operator:

# Create a local Kind cluster for development
kind create cluster --name lucid-dev

# Deploy your environment (mock TEE is auto-detected for local clusters)
lucid apply -f environment.yaml -y

Mock vs Production

Aspect Mock Mode Production
Attestation Software-simulated Hardware-backed
Security Development only Full security guarantees
AI Passport hardware_attested: false hardware_attested: true
Use case Local testing Production deployments

Never use Mock Mode in production

Mock mode provides no security guarantees. It exists solely for development and testing workflows.

Security Guarantees

What TEEs Protect Against

  • Malicious cloud operators: Even the cloud provider cannot read enclave memory
  • Compromised hypervisors: The hypervisor cannot access TEE data
  • Memory attacks: Enclave memory is encrypted in hardware
  • Tampering: Any code modification is detected via measurements

What TEEs Don't Protect Against

  • Side-channel attacks: Some timing/power analysis attacks may be possible (mitigated in newer hardware)
  • Bugs in enclave code: TEEs protect the environment, not the application logic
  • Denial of service: An attacker can prevent the enclave from running (but not read its data)

Further Reading