Identity federation and access to FMs

Lesson 2 of 3 in Enterprise Integration Architectures.

Access to a model is a chain, and the exam tests every link of it. Identity federation means the enterprise keeps its one identity source — Entra ID, Okta, Ping — and AWS trusts it: users and workloads exchange their IdP-issued identity (SAML or OIDC) for temporary AWS credentials, so no long-lived access keys exist for FM access at all. Role-based access control maps that identity to permissions at the group level: analysts assume a role that can invoke the summarization model; engineers assume one that can also manage Knowledge Bases; nobody gets person-by-person policies. Least privilege then shrinks each role to its real need — bedrock:InvokeModel on the specific model ARNs the application uses, not bedrock:* on *.

From corporate login to a scoped model call

  1. Enterprise IdP

    Entra ID, Okta, Ping — the single source of workforce identity, with its existing MFA, joiner/leaver, and group management.

  2. Federation (SAML/OIDC)

    IAM Identity Center for workforce access, or Amazon Cognito when a customer-facing app needs AWS credentials. Either way the IdP asserts who you are and which groups you hold.

  3. Assume mapped IAM role

    Group → role mapping is the RBAC layer: "genai-analysts" assumes the analyst role. Credentials are temporary and expire on their own.

  4. Least-privilege policy check

    The role allows bedrock:InvokeModel (and, deliberately, bedrock:InvokeModelWithResponseStream) on named model ARNs — and nothing else.

  5. Model answers

    The call travels via a VPC endpoint where network isolation is required; CloudTrail records who invoked what.

  6. Denied + logged

    Out-of-scope model, out-of-scope action, or expired session — the request dies at IAM, and the audit trail says so.

Data residency moves the question from who to where: some data may not leave a country, a regulated site, or even a building — yet the application still needs FM access. AWS Outposts extends AWS infrastructure, APIs, and tools onto your premises as a managed rack: the sensitive data is processed locally with the same programming model, and only what compliance permits crosses to the Region. AWS Wavelength embeds AWS compute at the edge of telecom 5G networks, so traffic from mobile devices reaches your application without ever traversing the public internet — the latency-and-locality play. Between site and Region, secure routing (Direct Connect or Site-to-Site VPN, plus VPC endpoints so Bedrock calls never touch public internet paths) carries whatever is allowed to travel.

Where the workload lives when jurisdiction constrains it
OptionWhat it isThe jurisdiction storyExam cue

AWS Region

Standard cloud deployment

Pick the Region inside the required jurisdiction; keep traffic private with VPC endpoints

"Data must remain in country X" where X has a Region

Outposts

AWS-managed racks/servers on your premises, operated as part of a Region

Local processing and local data residency for on-premises systems — same AWS APIs

"Must be processed on premises", "cannot leave the facility", "integrate with on-prem data"

Wavelength

AWS compute embedded in a telecom carrier's 5G network

Mobile traffic stays inside the carrier network — ultra-low latency at the mobile edge

"5G devices", "single-digit-millisecond", "edge deployment for mobile users"

Key terms: Identity federation, AWS Outposts, AWS Wavelength, GenAI gateway

Interactive checkpoint quiz (3 questions) — open this page in a browser to take it.