Skip to main content

Don’t Trust, Verify

Traditional AI APIs require you to trust that the provider handles your data securely. With confidential compute, you can verify these guarantees cryptographically:
  • Hardware attestation proves your request ran on genuine, secure GPU hardware
  • Response signatures prove the output came from the attested machine
  • Nonce verification proves the attestation is fresh, not replayed

How It Works

  1. Request attestation with a unique nonce you generate
  2. io.net routes your request to a confidential compute-enabled GPU machine
  3. GPU TEE generates a hardware attestation report with a signing key
  4. Verify the attestation report proves the machine is genuine and secure
  5. Run inference - responses are signed with the attested key
  6. Verify signatures to prove responses came from the attested machine

Key Components

Trusted Execution Environment (TEE)

The GPU runs inside a Trusted Execution Environment that provides:
  • Memory isolation - your data is encrypted in memory and inaccessible to the host
  • Code integrity - only authorized code can run inside the TEE
  • Hardware attestation - the GPU can prove its identity and configuration

Attestation Agent (Open Source)

The attestation agent running on GPU machines is fully open source. You can audit the code that generates attestation reports and signs responses: Repository: https://github.com/ionet-official/cc-attestation-agent-api This transparency allows you to:
  • Verify what code is running inside the TEE
  • Understand exactly what is being attested
  • Build confidence in the verification process

Attestation Reports

When you request attestation, you receive:
FieldDescription
gpuNVIDIA GPU attestation report proving the GPU identity and TEE state
cpuCPU attestation report (when available) for additional verification
image_digestSHA256 hash of the container image running in the TEE
signing_addressPublic key the machine will use to sign inference responses
nonceYour nonce echoed back, proving freshness

Response Signatures

Every inference response includes cryptographic signatures in the response headers:
HeaderDescription
textThe content that was signed
signatureCryptographic signature of the text
signing_addressPublic key that signed (matches attestation report)
signing_algoAlgorithm used for signing

What You Can Prove

CheckWhat It Proves
GPU attestation report is validResponse came from genuine NVIDIA GPU in TEE mode
image_digest matches releaseRunning container hasn’t been tampered with
signing_address matches attestationResponses are signed by the attested machine
Signature verifiesResponse was not tampered with in transit and signed on attested machine
Nonce matches your requestAttestation is fresh, not replayed

Privacy Guarantees

Confidential compute operates in Zero Data Retention (ZDR) mode:
  • Your prompts and responses are never stored
  • Only token counts are recorded for billing
  • No logs of conversation content exist

What’s Next