Day 2 - The 4 Core Entities of IAM: User, Group, Role, Policy

User: A Person, or a System That Acts Like One

An IAM User is an entity with long-term credentials. It signs in to the console with a login password, or authenticates to the CLI/SDK with an Access Key ID + Secret Access Key. The fact that these credentials never expire is both the biggest risk and the decisive point of divergence from IAM Roles.

🔍 Going deeper: For IAM User access keys, IAM Access Analyzer tracks "last used" information. Check the last-used time with aws iam get-access-key-last-used --access-key-id AKIA..., and deactivating keys unused for 90+ days is a security best practice. AWS's own security guide (IAM Best Practices) strongly recommends moving human credentials off IAM Users and onto AWS IAM Identity Center (formerly SSO). SSO integrates with ID providers (Okta, Azure AD, etc.) via SAML 2.0 or OIDC, and ultimately issues temporary credentials through STS, so long-term keys never exist at all.

📚 Case study: The Code Spaces incident of June 2014. Code Spaces, a GitHub code hosting company, had its AWS root account credentials stolen; its EC2, S3, EBS, AMIs, and snapshots were all deleted, and the company went out of business within 18 hours. The direct cause was no MFA + reliance on a single credential. Immediately after this incident, AWS began recommending safeguards far more aggressively, such as enforcing MFA on root accounts and the policy variable to require MFA for IAM Users (aws:MultiFactorAuthPresent).

Group: The Unit of Permission Bundling

An IAM Group is a container for grouping people, not a subject of permissions. Attach a policy to a Group and the permissions propagate to all Users in that group. A Group itself cannot call resources (a Group cannot be a Principal).

[ Engineering Group ]
    ├── Alice (User)
    ├── Bob (User)
    └── Policy: AmazonS3ReadOnlyAccess
                 ↓
        Both Alice and Bob can read S3

The core value of Groups is permission reuse and auditability. If you attach policies individually to 50 engineers, tracking who has what permission becomes impossible. With Groups, "the permissions of the Engineering Group = Alice's permissions", making change tracking easy. Note that Groups cannot be nested, and a User can belong to at most 10 Groups. The simple structure is intentional.

Role: The Producer of Temporary Credentials

The IAM Role is IAM's most powerful weapon and the concept most frequently tested. A Role is a bundle of permissions owned by no one, and when someone "assumes" the Role, STS issues temporary credentials. Temporary credentials consist of a 3-piece set — (1) Access Key ID, (2) Secret Access Key, (3) Session Token (this is the key part) — and expire automatically after 1 hour by default.

1. Lambda function starts
2. Lambda runtime calls STS AssumeRole
3. STS issues temporary credentials (valid for 1 hour)
4. Lambda code calls DynamoDB with those credentials
5. When they expire after 1 hour, the SDK automatically re-issues them

A Role attached to an EC2 instance behaves more subtly. EC2 exposes credentials through the IMDS (Instance Metadata Service). When the SDK GETs 169.254.169.254/latest/meta-data/iam/security-credentials/<RoleName>, temporary credentials are returned as JSON. These credentials are valid for about 6 hours, but the SDK automatically refreshes them 15 minutes before expiration.

🔍 Going deeper: The internal workings of AssumeRole are as follows. (1) A Principal calls the sts:AssumeRole API, specifying a RoleArn. (2) STS evaluates the Role's Trust Policy (which defines who can assume this Role). (3) If it passes, the Role's Permission Policy (what this Role can do) is granted. (4) STS issues temporary credentials, signing the permissions and expiration time into a JWT-like Session Token. (5) Subsequent API calls are signed with Signature V4, and AWS decodes the Session Token to verify permissions. The fact that both the Trust Policy and the Permission Policy must pass is the part that frequently confuses people.

💡 Memorization tip: A User is "a person with long-term credentials"; a Role is "a bundle of permissions that issues temporary credentials". If an exam scenario mentions "an EC2 instance accessing S3", the answer is a Role 99% of the time. If "hardcoding keys in code" appears, it is never the correct answer.

Policy: The Specification of Permissions Expressed in JSON

A Policy is a document that writes down permissions in JSON. Five fields are central: Effect / Action / Resource / Condition / Principal.

This policy means "allow GET/PUT on the S3 prefix corresponding to one's own username, only when MFA-authenticated and the request targets the Seoul region". ${aws:username} is a policy variable that is substituted with the caller's username at evaluation time. This is the core mechanism of ABAC.

🔍 Going deeper: AWS IAM evaluation logic has a clear priority order. (1) If there is an Explicit Deny, it is unconditionally denied (the strongest). (2) If the SCP (Organizations level) has no Allow, deny. (3) If a Permission Boundary exists and contains no Allow, deny. (4) An Allow must exist in an Identity-based or Resource-based policy for the request to be allowed. In other words, the key asymmetry is: "Allows accumulate as a union, but a Deny anywhere blocks immediately". Because of this asymmetry, SCPs are well suited to "making things absolutely impossible" (e.g., blocking root user actions, prohibiting the use of certain regions).

⚠️ Trap: A policy containing both "Resource": "*" and "Action": "*" is almost always AdministratorAccess-level. Developers commonly slap such policies into CloudFormation templates or SAM definitions "to quickly fix a permission issue", and IAM Access Analyzer detects this and sends alerts. On the exam too, when the keyword "least privilege" appears, a * policy is almost always a wrong answer.

📚 Case study: In 2017, NICE Systems, a partner of Verizon's DCS (a data integration company), exposed the call records of 14 million people through a publicly configured S3 bucket. The bucket policy contained both "Principal": "*" and "Effect": "Allow". After this incident, AWS added a series of safeguards: (1) Block Public Access enabled by default on all new buckets, (2) red warning badges on public buckets in the S3 console, (3) IAM Access Analyzer (launched December 2019) automatically detecting resources exposed externally.

Identity-based vs Resource-based: The Same Permission from Two Directions

The same "Alice can PutObject to my-bucket" can be expressed in two ways.

The difference shows up in cross-account scenarios. Within the same account, either one is sufficient, but when granting permissions to Alice in another account, both must be present. Our account's bucket policy must allow Alice, and Alice's account must also allow Alice's User to access our bucket. With only one side, the cross-account request is denied.

S3, KMS, SQS, SNS, Lambda, Secrets Manager, ECR, and others support resource-based policies. Services like DynamoDB, EBS, and CloudWatch do not, so cross-account access must always go through an IAM Role.

Permission Boundary and SCP: The Upper Limits of Permissions

These are two safeguards frequently used in large organizations. A Permission Boundary sets the upper limit of an IAM User/Role's permissions. Even if AdministratorAccess is attached via an identity-based policy, if the Permission Boundary is ReadOnlyAccess, the actual effect is ReadOnly.

An SCP (Service Control Policy) sets an absolute upper limit at the OU/Account level of AWS Organizations. If an SCP says "deny regions other than us-east-1 and us-west-2", then no user/role inside that OU can operate in any other region. SCPs do not grant permissions, they only restrict (think of them precisely as a deny-only filter).

💡 Memorization tip: The SCP is the company-wide constitution, the Permission Boundary is an individual's ID card, and the identity-based policy is the actual activity permission. If any one of them blocks, the action is impossible.

Simulating Permissions with the CLI

aws iam simulate-principal-policy rarely appears on the exam, but in practice it is gold for debugging.

The response contains EvalDecision: allowed | explicitDeny | implicitDeny. implicitDeny means "no explicit Allow was granted anywhere", and explicitDeny means "there is a Deny somewhere". For the latter, you should examine the SCP, the Permission Boundary, or the Resource Policy.

Wrapping Up

The four parts we saw today are the ABCs of the cloud permission model. A User is a person's long-term credentials, a Group is the unit of permission bundling, a Role is a permission vault that anyone can assume to receive temporary credentials, and a Policy is the specification of permissions expressed in JSON. These four are the inputs to the IAM evaluation engine, and the engine evaluates in the order SCP → Permission Boundary → Identity-based → Resource-based → Session Policy to produce a result.

In the next article, we go deep into what sits on top of this: STS (the temporary credential issuer), policy conditions (Condition), and the cross-account patterns of resource-based policies. Recalling that the Capital One incident was a case of IAM Role credentials leaking through IMDSv1, it becomes clear that advanced IAM is not mere theory but the key to preventing real incidents.