Technical Guide

Runtime Attestation for AI Models

Runtime attestation is a technical control that continuously verifies, using cryptographic evidence, that a deployed model, its configuration, and its execution environment match an approved reference state. It differs from static model validation, which checks a model once before deployment, and from monitoring or observability, which describes behavior rather than proving integrity.

Defining Runtime Attestation

Runtime attestation extends established trusted computing concepts to AI workloads. The IETF Remote ATtestation ProcedureS (RATS) architecture defines a general model built on three roles: an attester that generates evidence about a system's state, a verifier that evaluates that evidence against known-good reference values, and a relying party that acts on the result. Applied to AI, this means an attestation mechanism produces cryptographically signed evidence describing the model artifact, its configuration, and the environment it runs in, and a verifier checks that evidence before the system is trusted to operate.

This is distinct from testing a model's accuracy or safety before deployment. Static validation answers whether a model behaved correctly at a specific point in time under test conditions. Runtime attestation answers whether the model and environment currently running are the same ones that passed validation, and whether they remain unaltered during execution.

What Runtime Attestation Verifies

In practice, attestation evidence is intended to cover four related layers of the deployed system, each verified continuously rather than once.

Model Artifact

Cryptographic hash or signature confirming the deployed model matches the approved version.

Configuration

Verification that runtime parameters and settings have not been altered post-approval.

Execution Environment

Evidence that the container, host, or enclave running the model is unmodified and trusted.

Continuity

Repeated or ongoing verification, not a single point-in-time check.

Technical Mechanisms

Current implementations of runtime attestation for AI workloads rely on mechanisms adapted from general trusted computing and confidential computing, since no AI-specific attestation standard currently exists. Hardware-based Trusted Execution Environments, such as those built on Intel SGX/TDX or AMD SEV-SNP, isolate model weights and inference computation from the host operating system and hypervisor. These environments support remote attestation, allowing a verifier to confirm that code is running in a genuine, unmodified enclave before sensitive operations, such as releasing decryption keys for model weights, are permitted.

Cloud providers offer attestation services that implement this pattern. Microsoft Azure Attestation validates evidence from a platform or TEE against configured policies before granting access to protected resources. Google Cloud's confidential computing documentation describes a similar pattern of confirming genuine, unmodified execution before secrets are released. These services are general infrastructure capabilities rather than AI-specific products, but they form usable building blocks for attesting inference environments.

At the supply chain level, frameworks such as SLSA and in-toto provide cryptographically signed provenance for build artifacts. These have been referenced in ML security discussions as applicable to verifying that a training or deployment pipeline produced the model artifact being attested, connecting attestation at runtime back to a verifiable build history.

Implementation Considerations

Before runtime attestation can function, organizations need a signing and provenance process for approved model artifacts, consistent with SLSA and in-toto style practices. Without a trusted baseline to compare against, attestation evidence has nothing meaningful to verify.

Hardware-based attestation through TEEs may constrain deployment options, since it typically requires infrastructure with confidential computing capability. This is a deployment decision with cost and portability implications, not a universal requirement, and organizations should weigh it against software-only measurement approaches, which offer weaker trust guarantees but broader compatibility.

Attestation also introduces verification latency. Continuous or periodic re-attestation is architecturally different from a one-time boot-time check, and is necessary to detect drift introduced after initial deployment, such as configuration changes or unauthorized model swaps. The frequency of re-attestation needs to be balanced against performance requirements, particularly for real-time inference or agent action loops where added latency has direct operational cost.

Finally, existing AI observability and monitoring tools do not natively produce cryptographic attestation evidence. Attestation typically requires an additional tooling layer on top of existing MLOps and inference infrastructure, and key management for signing and verification needs to integrate with existing enterprise PKI or secrets management systems.

Governance and Accountability

NIST's AI Risk Management Framework and OWASP's Top 10 for LLM Applications both call for supply chain integrity and continuous verification, but neither prescribes a specific runtime attestation mechanism. This leaves organizations to determine their own approach, informed by general trusted computing standards rather than an AI-specific one.

No confirmed regulatory requirement currently mandates runtime attestation specifically for AI models. Existing regulatory attention tends to focus on broader risk management and documentation obligations rather than prescribed technical controls. That said, attestation evidence and logs are likely to become part of the audit and retention trail organizations maintain for AI governance purposes, since they provide a verifiable record distinct from behavioral logs.

Governance models should treat accountability for pre-deployment approval, such as model validation and testing, as distinct from accountability for runtime integrity assurance. These are different controls with different owners, and conflating them tends to create gaps where a model can drift from its approved state without detection between validation checkpoints.

Where This Fits Alongside Runtime Governance

Runtime attestation addresses one specific question: is the model and its environment unaltered from an approved state. It does not by itself enforce what an agent is permitted to do, verify agent identity, or govern which tools an agent can call. Those are separate but related concerns that sit within a broader runtime governance architecture, alongside agent identity, least-privilege permissions, tool approval workflows, and audit logging.

Trussed AI provides runtime governance and security for enterprise AI agents, covering areas including runtime policy enforcement, runtime monitoring, agent identity, agent permissions, and audit logging. Attestation evidence, where an organization has implemented it, is one input that a runtime policy engine can incorporate alongside identity and permission checks when deciding whether an agent action should be authorized.

Where Attestation Fits in AI Infrastructure

Attestation does not operate as a standalone check. It connects to several adjacent systems that together determine whether a verified model or agent is trusted to act.

  1. 1

    Agent or Workload Identity

    Attestation evidence needs to be bound to a verifiable identity rather than evaluated in isolation, so that a verified integrity state can be tied to a specific model instance or agent.

  2. 2

    Policy Engines

    A policy engine evaluates attestation evidence against approved reference values, such as an expected model hash or signed configuration, and determines whether execution should continue or be blocked.

  3. 3

    Tool-Call or API Gateways

    For AI agents that take actions, gateways handling tool calls may need to check attestation status as a precondition before authorizing sensitive operations.

  4. 4

    Reference Value Management

    Verifiers require a secure, updatable source of known-good values, meaning the provisioning and rotation of approved model hashes and configurations is itself a governed process.

Evaluation Questions for Security Engineers

Use these questions to assess whether a proposed or existing attestation mechanism actually addresses your integrity assurance needs.

  • What specific evidence does the mechanism verify: model weights, configuration, container image, or the full execution environment?
  • Does the approach rely on hardware roots of trust (TEE/TPM) or software-only measurement, and what trust assumptions does that introduce?
  • How does the system integrate with existing identity, policy enforcement, and monitoring tools rather than duplicating their function?
  • What is the re-attestation frequency, and what latency does it add to production inference or agent workflows?
  • How are known-good reference values generated, stored, and updated securely across the model lifecycle?

Strengthen Runtime Governance for AI Agents

Attestation verifies integrity. Runtime governance determines what happens next. See how policy enforcement, identity, and audit logging work together to secure AI agents in production.

Explore Runtime Governance