Industry leaders and standards groups unveiled a draft specification this week designed to give enterprises cryptographic proof that large language models (LLMs) and sensitive inference workloads ran exactly as contracted inside hardware-protected environments.

What the draft 'Confidential AI Attestation' spec covers

The proposal, published by the Confidential Computing Consortium (CCC) in collaboration with an industry working group of cloud providers, hardware vendors and enterprise customers, defines a standardized attestation token for AI workloads. The token binds four elements:

  • Cryptographic identity of the model binary or weights (a model hash and signed manifest);
  • Runtime environment attributes, including enclave type and firmware/TEE measurements;
  • Data and policy provenance describing allowed data sources and retention rules; and
  • Execution evidence such as a nonce-bound transcript or summarized telemetry that can be verified by an auditor.

The spec is explicitly designed for two common enterprise deployment patterns: private models hosted in cloud Confidential VMs or Nitro-style enclaves, and vendor-hosted LLM inference services that must supply provable evidence of runtime behavior to customers and auditors.

Why enterprises asked for this

As enterprises increasingly deploy LLMs for regulated use cases—customer service, loan underwriting, and clinical decision support—compliance teams seek auditable, tamper-evident proof that (a) the correct model was used, (b) protected data never left a trustworthy enclave, and (c) runtime policies (data retention, logging limits, prompt redaction) were enforced.

Until now, companies relied on contractual assurances, operator audits and provider dashboards. The new attestation approach aims to convert those assurances into cryptographically verifiable tokens that procurement, legal and security teams can ingest into governance systems.

Concrete enterprise benefits

  • Faster procurement: a standard attestation token can be included in SLAs and verified automatically during vendor selection.
  • Reduced audit burden: auditors can verify tokens rather than re-running invasive penetration tests on provider infrastructure.
  • Stronger incident response: tokens tied to execution timestamps and nonces help reconstruct what model/version executed for a given request.

How it works with existing tech

The spec builds on existing confidential computing primitives and remote attestation flows already supported by hardware vendors (Intel SGX, AMD SEV, and ARM Confidential Compute) and cloud services (confidential VMs, Nitro enclaves). It defines JSON Web Token (JWT)-style wrappers containing signed claims about model identity, enclave measurements and a pointer to an auditable telemetry bundle.

Crucially, the draft includes an interoperability section that maps vendor-specific attestation outputs into the standardized token format. That step is intended to prevent the ecosystem from fragmenting into provider-specific proofs — a major concern for enterprises that want to avoid vendor lock-in.

Industry reaction and vendor commitments

Several major cloud providers signaled support for the approach in parallel communications to enterprise customers and consortium members. Hardware vendors welcomed the work as a way to drive adoption of TEEs for higher-value AI workloads. Security and governance tooling vendors said they will add verification libraries and policy templates that consume the new tokens.

Startups that specialize in model provenance and AI supply-chain verification described the spec as a missing piece for end-to-end model governance: “Standardized runtime attestations close the loop between model SBOMs and auditor evidence,” one vendor briefing to the consortium argued.

Open questions and challenges

While the proposal targets a practical interoperability gap, several technical and operational challenges remain:

  1. Trust anchors and root-of-trust choices: who operates the global trust registry of attestation public keys, and how are revocations handled?
  2. Privacy of attestation data: balancing the need for auditability with leakage risks in telemetry summaries.
  3. Performance and cost: running high-throughput inference inside TEEs can incur latency and higher instance pricing.
  4. Key and identity management across multi-cloud and hybrid environments.

Consortium leaders acknowledged these trade-offs in accompanying guidance and proposed phased adoption: start with signing model manifests and enclave measurements, then add optional runtime telemetry for high-risk regulated workloads.

Next steps and timeline

The CCC opened the draft for a 60-day public comment period and is soliciting input from enterprise purchasers, auditors, cloud providers and regulators. The working group said it plans a revision cycle informed by feedback and a reference implementation that demonstrates cross-provider verification by Q4 2026.

For enterprises, the specification provides a concrete procurement lever to demand cryptographic proof of runtime compliance. For cloud providers and hardware vendors, it represents a route to differentiate secure AI hosting offerings while addressing customer demands around model governance and regulatory readiness.

What to watch

  • Which cloud providers publish SDKs and verification endpoints compatible with the spec;
  • Whether major enterprise buyers (banks, insurers, healthcare systems) require attestation tokens in new AI contracts;
  • How regulators reference attestation capabilities in upcoming AI oversight guidance and audits.

The Confidential AI Attestation draft is a practical step toward closing a long-standing enterprise gap: turning trust-in-contracts into trust-by-cryptography. If adopted broadly, it could become a foundational control in the next wave of enterprise LLM governance and procurement.