Implementation Guide

    Using AWS SCPs and Google Cloud Organization Policies to Force Model Traffic Through a Policy Proxy

    Make the proxy the only permitted route by denying direct model invocation and unapproved egress for every workload identity at the organization layer, then exempting only the proxy's own service identity. Add network controls so alternate paths do not exist, test for bypass after each change, and treat the proxy as the layer that evaluates each request, because org-level policy governs who can call which services, not what is inside a call.

    The pattern: four layers, one permitted route

    1. 1

      Organization-level deny

      An SCP in AWS and an Organization Policy in Google Cloud deny model invocation and unapproved service use for workload identities across member accounts and projects.

    2. 2

      Proxy identity exception

      The deny carries a narrow exception keyed to the proxy's role or service account, so the proxy can call model endpoints and nothing else can.

    3. 3

      Network containment

      Private endpoints, endpoint policies, Private Service Connect, VPC Service Controls as a perimeter, and egress controls limit which destinations workloads can reach.

    4. 4

      Per-request enforcement

      The proxy evaluates each request and applies a decision before it reaches the model.

    Bypass-testing checklist

    • A direct call to a cloud-native model service using the workload's own credentials is denied.
    • A direct call to a third-party or Azure-hosted model API from a workload network fails at egress.
    • The same calls succeed when routed through the proxy and are evaluated by it.
    • A workload cannot create an internet gateway, peering connection, or other alternate route.
    • A workload cannot assume or impersonate the proxy identity.
    • A newly created account or project inherits the deny without manual steps.

    Implementing the controls in AWS and Google Cloud

  1. Inventory model routes: native services, private endpoints, internet egress, third-party APIs.
  2. Attach organization-level deny policies at the OU or folder level.
  3. Exempt only the proxy identity, matched on identity rather than IP.
  4. Deny creation of alternate egress and routing constructs in member accounts.
  5. Contain network paths with private connectivity and a service perimeter.
  6. Roll out in stages, starting with non-production, and test each legitimate flow first.
  7. Why deploying a proxy does not make it mandatory

    A policy proxy enforces policy only on traffic that reaches it. If a workload holds native cloud credentials with permission to call Amazon Bedrock or Vertex AI, or has a network path to a third-party model API, it can skip the proxy without any technical resistance. Enforcement then depends on developers choosing the right endpoint, which is not a control.

    The gap affects two things. Runtime control is incomplete because some calls are never evaluated. Audit evidence is incomplete because the record of what AI was allowed to do covers only the traffic that happened to flow through the proxy. For a regulated enterprise, both gaps matter.

    This maps to the policy decision point and policy enforcement point model in NIST SP 800-207. Access is granted through a decision point and an enforcement point, and a gateway acting as a proxy allows only approved communication paths configured by the policy administrator. The organization layer is what makes the approved path the only path.

    Scoping the proxy exception so it does not become the bypass

    The exception for the proxy identity is the most sensitive part of the design. If it is broad, anyone who can assume or impersonate that identity inherits unrestricted model access, and the organization-level deny is meaningless.

    • Limit who can assume or impersonate the proxy identity, and treat changes to that permission as a reviewed event.
    • Grant the proxy identity only the model actions it needs, not blanket service access.
    • Keep the exception inside the policy statement itself so reviewers see the deny and its single exemption together.
    • Run separate proxy identities per environment so a development credential cannot reach production models.
    • Review the exception on a schedule and after any proxy architecture change.

    What org-level controls cannot do

    Organization-level policy decides which identities can call which services and which paths exist. It does not inspect the content of an individual request. It cannot redact sensitive data from a prompt, block a response that violates policy, or escalate a tool call for approval. Preventive controls also have scope boundaries, so confirm which accounts and identity types your organization-level policies actually cover and compensate with other controls where they do not apply.

    That is the job of enforcement in the execution path. Trussed enforces policy on individual interactions and tool calls, with allow, redact, and block decisions made at runtime rather than reviewed after the fact. It deploys self-managed in the customer cloud and generates audit-ready evidence from enforcement. The cloud controls described here ensure that traffic reaches that layer, and the proxy decides what happens to each interaction once it does.

    Treat the two as separate controls with separate owners and tests. The organization layer is tested by attempting to bypass it. The enforcement layer is tested by what it decides and records on traffic that arrives.

    Deny by default, allow through the proxy identity

    Organization layer

    AWS SCPs and Google Cloud Organization Policies deny direct model access for workload identities.

    Network layer

    Private connectivity and egress controls remove alternate paths to model endpoints.

    Proxy identity

    One service identity is permitted on the approved path, scoped narrowly.

    Execution path

    The proxy evaluates each interaction and records what it decided.

    Make the proxy the only route, then enforce on every interaction

    Talk to an expert about enforcing policy in the execution path in your own cloud and producing audit-ready evidence for a regulated enterprise.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.