Iam Best Practices

Identity & Access Management Best Practices

Least-privilege IAM design, key rotation, MFA enforcement, role assumption, and cross-account access patterns for AWS / GCP / Azure / Kubernetes

ALWAYS

• Grant the minimum permissions required for the workload's stated job (NIST AC-6). Start with a deny-by-default policy and add concrete actions with explicit Resource ARNs; never Action: "*" combined with Resource: "*".
• Prefer workload identity (IAM roles for service accounts on EKS, GKE Workload Identity, Azure Managed Identity) over long-lived access keys. Static access keys are the exception, not the default.
• Require MFA for every human IAM user, especially any principal that can assume a privileged role. Enforce MFA via an IAM policy condition (aws:MultiFactorAuthPresent: true), not just a directory-level setting.
• Rotate access keys, service-account keys, and signing keys on a documented schedule (≤ 90 days). Detect inactive credentials (≥ 90 days unused) and disable them automatically.
• Use role assumption with sts:AssumeRole + ExternalId for cross-account trust. The ExternalId must be unique per consumer and stored as a secret in both accounts.
• Issue session-scoped credentials with MaxSessionDuration ≤ 1h for human roles and ≤ 12h for break-glass roles. Long-lived sessions defeat rotation.
• Separate deploy and runtime identities. The CI/CD pipeline gets a deploy role; the running service gets a distinct runtime role with no IAM-mutating permissions.
• For Kubernetes RBAC, scope Role / RoleBinding to a single namespace; use ClusterRole only for true cluster-wide objects. Audit cluster-admin bindings on every PR.
• Log every IAM-mutating call (CloudTrail / Cloud Audit Logs / Azure Activity Log) to a tamper-evident sink. Alert on policy changes, iam:PassRole, iam:CreateAccessKey, and sts:AssumeRole from unexpected principals.
• For break-glass access (root, owner, cluster-admin), require an out-of-band approval (e.g., PagerDuty incident + ticket) and emit an immediate alert on every use.
• Tag every IAM principal with owner, environment, and purpose. Use these tags in SCPs / org policies to constrain blast radius.

NEVER

• Use the root account for day-to-day operations. Root credentials get a hardware MFA device, are stored offline, and are used only for the small set of root-only tasks (e.g., closing the account, changing the support plan).
• Embed long-lived access keys in source, container images, AMIs, or CI environment variables when a workload identity is available.
• Grant iam:PassRole with Resource: "*". Always pin the role ARNs the caller may pass to downstream services.
• Grant iam:* or sts:* to a runtime workload — these are deploy-time permissions only.
• Share a single IAM user across multiple humans or services. One principal per identity is the audit invariant.
• Use AdministratorAccess (or any *:*) managed policy on a routine basis; treat it as a break-glass-only attachment.
• Trust any cross-account assume-role without an ExternalId condition for third-party integrations (Confused Deputy: AWS Security Bulletin 2021).
• Hard-code AWS / GCP / Azure ARNs / resource IDs in policy documents without a corresponding tag-based or organization-path scope (when the number of resources can grow).
• Disable MFA for a principal to "fix" a login problem — rotate the device, do not remove the requirement.
• Persist OIDC / SAML assertion tokens beyond their stated TTL. Refresh by re-assertion, not by storing the original token.

KNOWN FALSE POSITIVES

• Resource: "*" is acceptable for inherently account-scoped read operations like ec2:DescribeRegions or sts:GetCallerIdentity — those APIs do not accept a resource ARN.
• Service-linked roles (e.g., AWSServiceRoleForAutoScaling) ship with broader permissions than your custom roles; that is by design and managed by the provider.
• One-time bootstrap operators (Terraform-runners in a fresh account) often need elevated permissions; gate by tag / SCP and revoke after the bootstrap is complete.
• Local-development emulators (LocalStack, GCS emulator) may accept any credentials; that is a property of the emulator, not a real grant.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ShieldNet-360
• Source: ShieldNet-360/secure-vibe
• License: MIT

Install and usage instructions live in the source repository linked above.