aimarket-tee

Documentation

TEE-attested execution for AI capabilities. AWS Nitro Enclaves / Intel TDX. Before invoke, the server sends a cryptographic attestation report: "this code runs in an encrypted hardware enclave. I physically cannot see your input." Receipts are signed by the enclave key. Unlocks enterprise, legal, medical, and finance verticals.

When to Use

• Confidential AI processing — consumer sends sensitive data (PII, legal docs, medical records), provider can't see it
• Enterprise compliance (GDPR/HIPAA/FedRAMP) — attestation report proves code integrity and data isolation
• Third-party model hosting — model owner deploys to enclave, hub operator can't extract weights
• Regulated industries — prove to auditors that execution happened in certified hardware

Installation

pip install aimarket-tee

API Endpoints

No additional API endpoints. Activated automatically via invoke flow — attestation is attached to every invocation response.

End-to-End Example

from aimarket_tee.tee_attestation import TEEAttestationService, EnclavePlatform

service = TEEAttestationService(platform=EnclavePlatform.AWS_NITRO)

# Execute capability inside enclave — provider sees only hashes
result = service.execute_with_attestation(
    capability_id="legal.review@v1",
    product_id="prod-legal",
    input_payload={"documents": {"nda": "CONFIDENTIAL: review this NDA..."}},
    code_identifier="legal-review-v2.0.0-githash-abc123",
    price_usd=5.00
)

print(result["attestation"]["platform"])  # aws_nitro
print(result["attestation"]["code_hash"])  # sha256 of code
print(result["receipt"]["input_hash"])     # sha256 of input (provider never sees plaintext)
print(result["enterprise_compliance"]["gdpr"])  # "Input never leaves enclave in plaintext"

Security

• Code hash verification — consumer verifies the exact code running in enclave
• Input confidentiality — provider sees only SHA-256(input), never plaintext
• Output signed by enclave key — verifiable proof of correct execution
• Attestation TTL — 5 minutes by default, prevents replay attacks

License

MIT · Maintained by AI-Factory