Audits

Confidential computing audits for AI startups

An independent review of how your AI product protects model weights, customer data, and inference traffic while they are in use, with a prioritized remediation plan.

Authorization comes first

CCSI audits only begin after your company requests one and provides written authorization from someone empowered to grant it, together with an agreed written scope (systems, accounts, environments, dates, and testing methods). We do not scan, probe, or test any organization’s systems without that authorization, and audit findings are confidential to the requesting company.

What is covered

Attestation and trust

Where and how attestation is verified
Measurement and policy management
Freshness and replay protection

Keys and secrets

Attestation-gated key release
KMS/HSM configuration and IAM
Secrets in images, logs, and environment

AI-specific risks

Model weight protection at rest and in use
Prompt and inference data handling
Isolation between tenants and pipelines

Supply chain

Build provenance and reproducibility
Base images and dependencies
Patch process for TEE and platform advisories

Operations

Debug and admin access paths
Logging without leaking data
Incident and TCB-recovery playbooks

Reporting

Findings rated by severity
Remediation roadmap
Optional path to CCSI business certification

Audit tracks

Pick the track that matches where your workloads run today. CCSI is independent and not affiliated with any cloud provider.

Track A

AWS confidential computing

For workloads using AWS Nitro Enclaves or other AWS confidential computing options. Reviews enclave image builds, attestation document verification, KMS key policies conditioned on attestation, vsock and parent-instance exposure, and IAM.

Track B

Google Cloud confidential computing

For workloads on Confidential VMs, Confidential GKE nodes, or Confidential Space. Reviews attestation token verification, workload identity and key release policies, image provenance, and operator access.

Track C

Not yet on confidential computing

For teams running on standard VMs, containers, or managed AI services. Maps where sensitive data is exposed in use, assesses current controls, and produces a migration plan to a suitable TEE platform.

Process

01
You submit an audit request
02
Scoping call; written scope and rules of engagement agreed
03
Signed written authorization received
04
Review and authorized testing within scope
05
Confidential report and remediation review

Pricing depends on scope: Contact us.