API Gateway Security Architecture (2026): WAF, Cognito, Private APIs, and Verified Permissions

In June 2026, Amazon Cognito launched multi-Region replication so user pools on Essentials or Plus can keep authentication available in a standby Region during a regional disruption (What’s New, developer guide). Amazon Verified Permissions continues to wire Cedar policies to API Gateway via a console wizard and Lambda authorizer for group- and attribute-based access (AVP + API Gateway, Cognito AVP authz).

This is the API security architecture — edge → authz → private vs public → abuse. It is not REST/HTTP/WebSocket patterns, not JWT vs introspection, and not a Shield Advanced buyer guide.

Artifacts: authz decision matrix, hardening checklist, architecture diagram (draw.io).

Benchmark silhouette (not a cited client) — B2B SaaS API, ~18k RPS peak, Cognito user pool, public REST + two private partner APIs. Regional auth drill before MRR: auth 5xx 4.2% for 14 minutes. After Cognito MRR + multi-Region OIDC issuer: drill 5xx 0.1%. Same month, a scraper hit an unauthenticated export route: ~$1.8k excess egress/compute until a WAF rate-based rule landed. Lesson: auth DR ≠ abuse protection.

Four layers — edge to backend

Opinionated take: Put WAF in front of every public API before you tune Cedar policies. Authorization elegance does not stop a cheap L7 scrape.

Reference architecture

Clients ──► CloudFront ──► AWS WAF ──► API Gateway (public)
                                      │
                                      ├── JWT authorizer (Cognito)
                                      │      └── optional AVP Lambda authorizer (Cedar)
                                      └── integration ──► Lambda / VPC Lattice / HTTP

Internal ──► Private API Gateway ◄── VPC endpoint + resource policy
                                      └── IAM SigV4

Private API patterns with VPC Lattice for cross-account exposure are documented in AWS Containers guidance (API Gateway + VPC Lattice).

Choose the authorizer

Use authz-decision-matrix.md:

• End users: Cognito JWT
• Fine-grained / multi-service rules: Verified Permissions
• AWS service callers: IAM
• Public read: WAF + cache; still inventory every method

What broke — JWT without rate limits

What broke — A partner integration test script rotated through 12k token refreshes/hour against /oauth2/token adjacent paths and a weakly protected export. Cognito and API Gateway returned valid 401/403 mixes; WAF had no rate-based rule on that stage. Detection: Cost anomaly + Count mode rule simulation the following Monday. Fix: rate-based rule + usage plan metering + move export behind AVP policy requiring partner.export action.

Cognito multi-Region — what it fixes

MRR replicates user directory and config to a secondary Region (max 2 replicas). Requirements include Essentials/Plus, a multi-Region CMK, and a multi-Region OIDC issuer. It does not replicate your API Gateway stages or data stores — pair with your API DR runbook.

What to Do This Week

1. Inventory every API method: public vs private, authorizer type, WAF association.
2. Score authz needs with the decision matrix; schedule AVP only where groups are insufficient.
3. Add rate-based WAF rules on auth, search, and export.
4. If auth is a single-Region SPOF, evaluate Cognito MRR prerequisites (plan tier, CMK, OIDC issuer).

What This Post Doesn’t Cover

• WebSocket fan-out security edge cases
• mTLS partner termination design
• Full Shield Advanced commercial decision (buyer guide)
• AppSync-specific auth modes