If your team is still using shared IAM user credentials to run Terraform, it's time to switch to AWS SSO (IAM Identity Center). In this article, I'll walk you through how I migrated our multi-account Terraform setup from a shared deployment IAM user to individual SSO-based authentication for both local development and CI/CD pipelines.
Our Previous Setup
We had a classic multi-account Terraform setup with three AWS accounts:
- Shared/management account - Hosted the S3 state bucket, DynamoDB lock table, and custom Terraform modules in S3
- Dev account - Development environment
-
Live/prod account - Production environment (with additional
live-euandlive-drworkspaces)
A single IAM user called deployment lived in the shared account. It had an access key that was shared across the team and stored as GitHub secrets for CI/CD. The Terraform provider used assume_role to switch into the target account:
provider "aws" {
region = "us-east-1"
assume_role {
role_arn = "arn:aws:iam::<account_id>:role/deployment"
session_name = "deployment"
}
}
The S3 backend stored state and locks in the shared account:
backend "s3" {
bucket = "my-tf-states"
region = "us-east-1"
key = "core.tfstate"
dynamodb_table = "terraform-locks"
}
And our GitHub Actions workflow used static IAM keys:
- name: Configure AWS Credentials
uses: aws-actions/configure-aws-credentials@v4
with:
aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
aws-region: us-east-1
Custom modules were stored in an S3 bucket and referenced like:
module "my_module" {
source = "s3::/my-tf-modules/1.0.41/my-module.zip"
}
This setup worked, but it had serious problems:
-
No individual accountability - CloudTrail logs showed
deploymentuser for every change, making it impossible to trace who did what - Security risk - Static keys can leak, get committed to git, or be shared insecurely
- Key rotation pain - Rotating one shared key means updating it everywhere
- No MFA enforcement - Long-lived access keys bypass MFA requirements
The Solution: AWS SSO + OIDC
With AWS IAM Identity Center (SSO), each DevOps engineer authenticates with their own identity. For CI/CD, GitHub Actions uses OIDC federation - no static keys stored as secrets.
Local Development:
Engineer -> AWS SSO Login -> Temporary Credentials -> Terraform
CI/CD (GitHub Actions):
GitHub Actions -> OIDC Token -> AWS STS -> Temporary Credentials -> Terraform
Step 1: Remove the assume_role Block
Since each engineer will authenticate directly via SSO into the target account, there's no need for assume_role. Terraform will use whatever credentials are in the environment.
Before:
provider "aws" {
region = "us-east-1"
assume_role {
role_arn = "arn:aws:iam::<account_id>:role/deployment"
session_name = "deployment"
}
}
After:
provider "aws" {
region = "us-east-1"
}
Step 2: Fix the S3 Backend for Cross-Account State Access
This was the trickiest part. Our state bucket and DynamoDB lock table lived in the shared account, but now we're authenticating directly into dev/live accounts via SSO.
The problem: Terraform's S3 backend always looks for the DynamoDB lock table in the caller's account. So when authenticated as dev account, it looks for the lock table in dev - not in the shared account where it actually exists.
The fix: Add a profile to the backend config pointing to the shared account:
backend "s3" {
bucket = "my-tf-states"
region = "us-east-1"
key = "core.tfstate"
dynamodb_table = "terraform-locks"
profile = "shared-account"
}
This ensures both S3 and DynamoDB calls go to the shared account, while the provider uses your SSO profile for the target account.
You'll also need an S3 bucket policy on the state bucket to allow cross-account access:
{
"Version": "2012-10-17",
"Statement": [
{
"Sid": "TerraformStateAccess",
"Effect": "Allow",
"Principal": {
"AWS": [
"arn:aws:iam::<dev_account_id>:root",
"arn:aws:iam::<live_account_id>:root"
]
},
"Action": [
"s3:ListBucket",
"s3:GetObject",
"s3:PutObject",
"s3:DeleteObject"
],
"Resource": [
"arn:aws:s3:::my-tf-states",
"arn:aws:s3:::my-tf-states/*"
]
}
]
}
After changing the backend config, reinitialize:
terraform init -reconfigure
Step 3: Set Up AWS SSO Profiles
Each engineer adds profiles to their ~/.aws/config - one per account:
# Dev account
[profile dev]
sso_start_url = https://your-org.awsapps.com/start/#/
sso_region = us-east-1
sso_account_id = <dev_account_id>
sso_role_name = SuperAdmin
region = us-east-1
# Production account
[profile prod]
sso_start_url = https://your-org.awsapps.com/start/#/
sso_region = us-east-1
sso_account_id = <live_account_id>
sso_role_name = SuperAdmin
region = us-east-1
# Shared account (for Terraform state backend)
[profile shared-account]
sso_start_url = https://your-org.awsapps.com/start/#/
sso_region = us-east-1
sso_account_id = <shared_account_id>
sso_role_name = SuperAdmin
region = us-east-1
Replace SuperAdmin with your SSO permission set name.
If you manage Linux servers alongside your Terraform infra, LinuxTools.app is a handy CLI reference I use daily.
Step 4: Run Terraform Locally
# Login to SSO (opens browser for authentication)
aws sso login --profile dev
aws sso login --profile shared-account
# IMPORTANT: Clear any old static credentials first
unset AWS_ACCESS_KEY_ID
unset AWS_SECRET_ACCESS_KEY
unset AWS_SESSION_TOKEN
# Set the profile for the target account
export AWS_PROFILE=dev
# Verify you're using the SSO role (not the old IAM user)
aws sts get-caller-identity
# Should show: arn:aws:sts::<dev_account_id>:assumed-role/AWSReservedSSO_SuperAdmin_.../you@company.com
# Run Terraform
terraform init
terraform workspace select dev
terraform plan
terraform apply
Gotcha: If
AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEYare set in your environment, they take precedence overAWS_PROFILE. Alwaysunsetthem first. I spent a while debugging this -aws sts get-caller-identitykept showing the olduser/deploymentidentity.
Step 5: Set Up GitHub Actions with OIDC
We already had an OIDC provider configured in AWS using the unfunco/oidc-github/aws module. If you don't have one yet, add it:
module "oidc_github" {
source = "unfunco/oidc-github/aws"
version = "1.8.0"
github_repositories = [
"your-org/your-terraform-repo"
]
attach_admin_policy = true
}
output "oidc_role_arn" {
value = module.oidc_github.iam_role_arn
}
Then update the GitHub Actions workflow:
Before (static keys):
permissions:
contents: read
steps:
- name: Configure AWS Credentials
uses: aws-actions/configure-aws-credentials@v4
with:
aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
aws-region: us-east-1
After (OIDC):
permissions:
id-token: write # Required for OIDC
contents: read
steps:
- name: Configure AWS Credentials
uses: aws-actions/configure-aws-credentials@v4
with:
role-to-assume: ${{ secrets.AWS_OIDC_ROLE_ARN }}
aws-region: us-east-1
Key changes:
- Added
id-token: writepermission (required for GitHub to issue OIDC tokens) - Replaced
aws-access-key-id/aws-secret-access-keywithrole-to-assume
Set AWS_OIDC_ROLE_ARN per GitHub environment:
- development environment: OIDC role ARN from your dev account
- production environment: OIDC role ARN from your prod account
Step 6: Upgrade AWS Provider and Modules
After switching to SSO, I hit this error on terraform plan:
An argument named "enable_classiclink" is not expected here.
This happened because we were on AWS provider 4.67 and VPC module 3.18.1. EC2-Classic was fully retired by AWS, and these older versions still reference deprecated classiclink attributes.
The fix:
# Provider: 4.67 -> 5.x
required_providers {
aws = {
source = "hashicorp/aws"
version = "~> 5.0"
}
}
# VPC module: 3.18.1 -> 5.16.0
module "vpc" {
source = "terraform-aws-modules/vpc/aws"
version = "5.16.0"
}
Then run:
terraform init -upgrade
Step 7: Migrate Module Sources from S3 to GitHub
We had custom Terraform modules stored in an S3 bucket:
module "my_module" {
source = "s3::/my-tf-modules/1.0.41/my-module.zip"
}
After switching to SSO, terraform init failed with:
NoCredentialProviders: no valid providers in chain
The root cause: Terraform's module downloader uses the Go AWS SDK internally, which does not respect AWS_PROFILE when downloading S3 sources. The cleanest fix was switching to GitHub as the module source:
module "my_module" {
source = "github.com/your-org/your-tf-modules//my-module?ref=v1.0.93"
}
Common Pitfalls
1. Old credentials overriding SSO
Environment variables take precedence over AWS_PROFILE. Always clear them first:
unset AWS_ACCESS_KEY_ID AWS_SECRET_ACCESS_KEY AWS_SESSION_TOKEN
2. DynamoDB lock table not found in the right account
Without profile in the backend config, you'll get:
AccessDeniedException: User is not authorized to perform: dynamodb:PutItem
3. S3 module downloads failing with NoCredentialProviders
Switch to GitHub-hosted modules or export credentials before terraform init:
eval "$(aws configure export-credentials --profile dev --format env)"
terraform init
4. Missing id-token: write permission in GitHub Actions
OIDC requires id-token: write permission. Without it, the OIDC token request fails silently.
5. Backend configuration changed error
After adding profile to the backend:
Error: Backend configuration changed
Run terraform init -reconfigure to fix.
Summary
| Component | Before | After |
|---|---|---|
| Local auth | Shared deployment IAM user + static keys |
Individual SSO login per engineer |
| CI/CD auth | Static keys in GitHub secrets | OIDC federation (no secrets needed) |
| Provider config |
assume_role to deployment role |
No assume_role (uses env credentials) |
| State backend | Direct access from shared account |
profile pointing to shared account |
| Audit trail | "deployment" user for everyone | Individual engineer identity in CloudTrail |
| Key rotation | Manual, painful, shared | Automatic, per-session, no keys to manage |
| Module sources | S3 bucket (breaks with SSO) | GitHub repo (uses git credentials) |
| AWS provider | 4.67 | ~> 5.0 |
| VPC module | 3.18.1 | 5.16.0 |
The migration took some troubleshooting, but the security benefits are significant. Every terraform apply is now traceable to an individual engineer in CloudTrail, credentials are short-lived and automatically rotated, and there are no static keys to leak or manage.
If you are working with Linux servers as part of your infrastructure, check out LinuxTools.app - a free reference for CLI commands and utilities I use daily.
Written by Khimananda Oli. Find more DevOps and cloud infrastructure content at khimananda.com.
Top comments (0)