Back to Field Notes
Identity & Access/Field Note

Modern IAM Architecture — A Reference for 2026

IAM is no longer one product. It is a stack of five with a defined boundary between each, an integration backbone that costs as much as the components, and a workload + agent identity tier that most programs still pretend does not exist.

Author

Yusuf Ahmed

Principal Engineer, Platform Security

Published

March 12, 2026

Read

13 min

Share
AI-generated illustration of a banking facility
AI-generated illustration of a banking facility
Key Takeaways
  • 01Modern IAM is five products with hard boundaries: IdP (workforce), CIAM (customer), PAM (privileged), IGA (governance), ITDR (runtime detection). Anything that mixes them produces audit nightmares.
  • 02The integration backbone — SCIM, event streams, identity-event bus — should be budgeted equal to the components. It is where the program leaks, not where the products fail.
  • 03ITDR is the 2024–2026 addition most programs are still missing. Token theft, OAuth-app abuse, and federation persistence go undetected without it.
  • 04Workload identity (SPIFFE / cloud IAM) and AI agent identity are now first-class. The IAM team must own the schema even if platform engineering operates the runtime.
  • 05One identity-of-record per actor. Joint accounts, shared service principals, and 'team' OAuth apps are the most common root causes of identity-driven breaches.

Identity and access management is no longer a single product purchase. The modern stack has clear boundaries between an enterprise identity provider (IdP), customer identity (CIAM), privileged access management (PAM), identity governance and administration (IGA), and identity threat detection and response (ITDR). Workload identity and AI agent identity now sit alongside as first-class peers.

This piece is the reference we hand a CISO who is consolidating an IAM estate inherited through M&A, and the engineering manager who has to design the integration backbone that makes the consolidation actually work.

The five-component reference

Each component has a hard boundary. Cross-boundary leakage — a customer record in the workforce IdP, a service principal managed in PAM but not IGA — is where audit findings and breaches both originate.

/COMMON_LEAK

The deprovisioning gap that bites every M&A integration.

An employee leaves. IGA fires the deprovisioning workflow. The IdP disables the account. But the PAM vault still holds standing credentials for that employee, the CIAM tenant for an integrated B2B customer still references the user, and the ITDR allow-lists still trust their device. Without an event-bus integration, you have created a textbook insider-becomes-outsider risk. Audit it quarterly.

/MODERN_IAM_STACK

ComponentOwnsDoes NOT own
IdPWorkforce auth, SSO, MFA, lifecycle for non-privilegedPrivileged elevation, customer auth
CIAMCustomer auth/authz, marketing consent, B2C lifecycleWorkforce identities; never share users
PAMPrivileged access, JIT elevation, session recordingStanding admin grants (the anti-pattern)
IGAJoiner-mover-leaver, certifications, SoD policyRuntime detection (that is ITDR)
ITDRRuntime identity attack detection + responseLifecycle decisions (those are IGA)

The integration backbone — SCIM and the identity event bus

Each component owns a piece of the truth. The integration backbone — typically SCIM for object propagation and a Kafka-style identity event bus for behavior — is what makes them coherent. Most programs underspend here by 5–10x relative to the components, and pay the price in operational drift.

/IDENTITY_EVENT_BUS · canonical event shape

# Every component publishes onto, and subscribes from,
# a shared identity event bus. Events are versioned, replayable,
# and tagged with a global subject identity-of-record.

{
  "event_id":   "evt_01HX9...",
  "event_type": "identity.user.deprovisioned",
  "version":    "1.2",
  "ts":         "2026-04-21T11:42:13Z",
  "actor": {
    "subject_id": "ior:workforce:u_8421",   # identity-of-record
    "tenant":     "alexa-corp",
    "kind":       "human"
  },
  "source": "iga.sailpoint",
  "data": {
    "reason":      "termination",
    "effective":   "2026-04-21T11:42:00Z",
    "downstream":  ["idp.entra", "pam.cyberark",
                    "ciam.b2b-tenant-7", "itdr.sentinel"]
  }
}

# Every named downstream MUST acknowledge within SLA (60s).
# Unacked = page; auto-quarantine the subject after 5 min.

ITDR — the addition most programs are missing

Identity Threat Detection and Response is the newest tier and the most often skipped. The bet ITDR makes — that identity attacks (token theft, OAuth-app consent abuse, federation persistence, MFA fatigue) are now the dominant breach path — is borne out by every public 2024–2025 breach we have studied. ITDR catches what SIEM correlation rules struggle to express.

  • 01Stolen-session-token detection — token used from a fingerprint mismatch with the issuance device
  • 02OAuth-app consent anomaly — new app gains broad scopes, used immediately at scale
  • 03Federation persistence — non-IdP-admin modifies SAML trust or adds a federation domain
  • 04MFA fatigue — burst of push notifications followed by a successful approve
  • 05Service-principal credential anomaly — secret rotated outside change window or used from unfamiliar geo

Workload + agent identity — the new first-class tier

By 2026, workload identities (services, CI runners, batch jobs) outnumber human identities by 30:1 in most enterprises. AI agent identities — where an agent acts on behalf of a user with delegated authority — add another category. The IAM team must own the schema and lifecycle for both, even if platform engineering operates the runtime.

/INSIGHT

The one rule that scales for non-human identity.

One identity-of-record per actor, with a named human owner, a defined lifecycle owner, and a maximum credential lifetime. Shared service accounts, ownerless OAuth apps, and 'team' SaaS integrations are the three patterns we see in nearly every identity-driven breach. The fix is structural: refuse to provision the next one.

/NON_HUMAN_IDENTITY_TIER

ClassAuthenticationLifecycle ownerCommon failure
Workload (cluster-internal)SPIFFE / SPIRE attestationPlatform Eng + IAMmTLS-only without identity attestation
Workload (cloud-native)Cloud IAM workload identity federationPlatform Eng + IAMLong-lived static keys
CI/CD runnerOIDC-to-cloud federationDevX + IAMShared secrets across pipelines
AI agent (delegated)User OIDC + agent constraint claimsAI Platform + IAMUnbounded scope; no time-limit
3rd-party SaaS-to-SaaSOAuth app + scoped grantApp Owner + IAMOver-broad scopes at consent time

A reference topology

/TOPOLOGY · enterprise IAM 2026

              ┌──────────────┐    ┌──────────────┐
              │  Workforce   │    │   Customer   │
              │     IdP      │    │     CIAM     │
              └──────┬───────┘    └──────┬───────┘
                     │  SCIM + events    │
                     ▼                   ▼
              ┌────────────────────────────────────┐
              │       IDENTITY EVENT BUS           │
              └──────┬─────────┬─────────┬─────────┘
                     │         │         │
                ┌────▼───┐ ┌───▼────┐ ┌──▼─────┐
                │  PAM   │ │  IGA   │ │  ITDR  │
                │ (JIT)  │ │ (life- │ │ (runtime
                │        │ │ cycle) │ │ detect)│
                └────────┘ └────────┘ └────────┘
                     │
                ┌────▼────────────────────────────┐
                │   Workload + Agent Identity     │
                │  SPIFFE · CloudIAM · OIDC fed   │
                └─────────────────────────────────┘

Four KPIs that prove the IAM program is working

/IAM_BOARD_KPIS

KPITargetWhat it proves
Lifecycle SLA (JML propagation)≥ 99% within 60 s across all componentsBackbone is real
Standing privilege countTrending toward zero quarter-over-quarterJIT is actually used
ITDR alert MTTR< 30 min for high severityRuntime detection has muscle
Non-human identity coverage100% with named owner + lifetimeWorkload tier is governed

A 12-month plan to land the reference

  1. 01Q1 — Stand up the identity event bus. Migrate first three downstream integrations to event-driven. Establish lifecycle SLA measurement.
  2. 02Q2 — Roll out PAM JIT for top 20% of standing-admin grants. Begin ITDR deployment with the five highest-yield detections (token theft, OAuth abuse, federation persistence, MFA fatigue, service-principal anomaly).
  3. 03Q3 — Workload identity inventory + SPIFFE pilot in one production tier. Eliminate 50% of long-lived cloud IAM static keys via federation.
  4. 04Q4 — AI agent identity schema. Brief board against the four KPIs.
/MONDAY_PLAYBOOK

Monday morning — the diagnostic that takes 10 minutes.

Pick a recent termination. Trace the deprovisioning event across all five components plus the workload identity tier. Time it. If any downstream took more than 60 seconds, OR if any required manual intervention, the integration backbone has gaps. Open one ticket per gap and fix the worst one this sprint.

  • ▸Pick one termination from the past 7 days
  • ▸Trace the event across IdP, CIAM, PAM, IGA, ITDR, workload tier
  • ▸Measure end-to-end propagation time
  • ▸Any > 60s or any manual step = backlog item

Closing — IAM is the substrate

Every other security investment — Zero Trust, SASE, AppSec, cloud security — assumes the IAM substrate is correct. It usually is not. Five components with hard boundaries, an integration backbone that costs as much as the components, ITDR for runtime, and a workload + agent identity tier that the IAM team owns the schema for. Build that, and every downstream control gets cheaper and more defensible.

#IAM#IdP#PAM#IGA#ITDR#Workload Identity

/WRITTEN_BY

Yusuf Ahmed

Principal Engineer, Platform Security · Alexa Cybersecurity