gebalamariusz/cloud-audit

Added

• Data Perimeter Scanner - new check category (aws-dp-001 .. aws-dp-005)
  that evaluates resource-based policies for the two boundary failures detectable
  from a single account's configuration - read-only, and without AWS Organizations
  access:

  • Confused deputy: an Allow statement grants an AWS service principal
    without an aws:SourceAccount / aws:SourceArn / aws:SourceOrgID /
    aws:SourceOwner condition (MEDIUM). A service acting on behalf of another
    account could be coerced into operating on your resource.
  • Cross-organization exposure: an Allow statement grants a wildcard
    ("*") or external-account AWS principal without an organization-boundary
    condition (aws:PrincipalOrgID, aws:ResourceOrgID, aws:SourceOrgID, ...).
    Wildcard grants are HIGH; a named external account is LOW (often an
    intentional partner share). On Secrets Manager these are escalated to
    CRITICAL (wildcard) / MEDIUM (external) as a direct credential-exfiltration
    path.

  Services covered: S3 bucket policies (aws-dp-001), SNS topic policies
  (aws-dp-002), SQS queue policies (aws-dp-003), Secrets Manager resource
  policies (aws-dp-004), and Lambda resource policies (aws-dp-005). Every
  finding ships CLI + Terraform remediation and a breach-cost estimate. Condition
  keys are matched case-insensitively, and condition values are evaluated (not
  just key presence) - a guardrail scoped to a foreign account or org is still
  flagged. Default SNS/SQS policies relying on aws:SourceOwner set to the owner
  account are correctly treated as account-scoped (no false positive). Federated
  provider ARNs in a foreign account are flagged as external; federated URL IdPs
  and CanonicalUser principals are out of scope.

  This closes a gap the dominant open-source scanner explicitly does not cover
  (Prowler issue #7114, "Integrating SCP/RCP Policy Awareness", open since
  2025-03): evaluating data-perimeter condition keys on resource policies. SCP/RCP
  enforcement at the AWS Organizations level is intentionally out of scope.
  Detection taxonomy follows the AWS data perimeter whitepaper and the AWS
  cross-service confused-deputy guidance.

• Proof Mode - new opt-in scan --verify that cross-checks each detected IAM
  privilege-escalation path against the read-only iam:SimulatePrincipalPolicy
  API to confirm the principal's policies actually allow the required actions. Each
  path is annotated verified with verification_detail evidence: true = the
  simulator allowed every required action (policy-allowed; simulated, not
  executed); false = the simulator denied one (the static path is likely a
  false positive); null = not asserted. The scan prints
  Proof Mode: N/M escalation path(s) policy-allowed by IAM simulator. Honest
  framing throughout: this confirms the permission exists, not end-to-end
  exploitability - the simulator does not factor in SCPs, permission boundaries,
  resource policies, or trust conditions. Resource-scoped methods (iam:PassRole,
  sts:AssumeRole, compute-hijack) are deliberately left null because without
  resource ARNs the simulator evaluates against * and would over-report; paths
  gated by unevaluated condition keys are also left null.
  iam:SimulatePrincipalPolicy has no per-call charge, so the opt-in controls
  latency/throttling only; calls are deduplicated per (principal, action-set). New
  module proof.py; EscalationPath gains verified / verification_detail
  (backward compatible, default unchecked). Brings the 2026 "proof, not
  probability" validation trend - which commercial scanners sell as a paid
  flagship - to open-source AWS.

• AgentCore checks - first OSS scanner with dedicated Amazon Bedrock AgentCore
  (GA 2025-10) coverage: 6 read-only checks aws-agc-001..006 over the AI agent
  platform. Flags Code Interpreter / Runtime in PUBLIC network mode (egress
  exfiltration), Runtime not enforcing MMDSv2 (metadata credential theft), Memory
  without a customer-managed KMS key, Gateway with no inbound authorizer
  (authorizerType=NONE), and Gateway with no enforcing policy engine (missing or
  LOG_ONLY). Each finding ships a CLI + Terraform fix and breach-cost estimate,
  and maps to Palo Alto Unit 42 "Cracks in the Bedrock" research. Read-only
  bedrock-agentcore-control (no per-call charge); the service is regional and
  absent regions are skipped silently. New module agentcore.py; field names,
  operations and enum values verified against the live boto3 service model.

Changed

• Check count: 99 -> 110 (across 25 services). AgentCore adds a new service module
  of 6 checks; the data perimeter's 5 checks remain a cross-cutting category over
  existing services, like Threat Feed.

Tests

• 836 -> 948 (+112). New file tests/aws/test_data_perimeter.py: 43 unit tests
  pinning the _find_perimeter_gaps detection engine (confused deputy, cross-org
  wildcard/external, condition-value evaluation so a guardrail pointing at a
  foreign account/org is still flagged, federated provider ARNs, case-insensitive
  conditions, Deny/NotPrincipal/CanonicalUser/malformed handling) plus 17 moto
  integration tests across all five services. New file tests/test_proof.py:
  28 tests pinning the Proof Mode engine (decision semantics: allowed /
  explicit+implicit deny / unknown-decision / incomplete / empty / malformed; the
  resource-scoping gate that leaves PassRole/AssumeRole and deny-removal paths
  unasserted; condition-key gating; API-call dedup; per-path error isolation;
  provider-backed path). New file tests/aws/test_agentcore.py: 24 fixture-based
  tests (moto lacks bedrock-agentcore-control) covering each check's positive and
  negative cases across network mode, MMDSv2, memory KMS, gateway authorizer and
  policy engine, empty policy config, multi-page pagination, plus
  region-unavailable / access-denied skip and multi-region aggregation.

Compliance

• aws-dp-001 .. aws-dp-005 mapped: SOC 2 (CC5.2, CC6.1, CC6.6, C1.1),
  CIS AWS v3.0 (2.1.4), HIPAA (164.308(a)(1)(ii)(A), 164.312(a)(1)),
  ISO/IEC 27001:2022 (A.8.3, A.8.12), NIS2 (NIS2-RM-09a), BSI C5:2020 (IDM-07).