prowler-compliance

Imported from professor-moody/cloud-tools (bin/Prowler/skills/prowler-compliance/SKILL.md). Install upstream with npx skills add professor-moody/cloud-tools --skill prowler-compliance. Copyright stays with the author (Apache-2.0).

When to Use

Use this skill when:

• Creating a new compliance framework for any provider — decide universal vs legacy first (see below)
• Syncing an existing framework with an upstream source of truth (CIS, FINOS CCC, CSA CCM, NIST, ENS, etc.)
• Adding requirements to existing frameworks, or extending a universal framework to a new provider
• Mapping checks to compliance controls
• Adding ConfigRequirements guardrails so configurable checks can't silently satisfy a requirement with a loosened config
• Auditing existing check mappings as a cloud auditor ("are these mappings correct?", "which checks apply?", "review the mappings")
• Adding a new legacy output formatter (table dispatcher + per-provider classes + CSV models)
• Fixing JSON bugs: duplicate IDs, empty Version, wrong Section, stale check refs, inconsistent FamilyName, padded tangential check mappings
• Investigating why a finding/check isn't showing under the expected compliance framework in the UI
• Understanding compliance framework structures and attributes

The authoritative contributor doc is docs/developer-guide/security-compliance-framework.mdx — keep this skill and that doc consistent when either changes. For reviewing a compliance PR, use the sister skill prowler-compliance-review instead.

Universal vs Legacy: The First Decision

Prowler supports two JSON schemas. Choosing wrong means unnecessary Python code, so decide this before anything else. At load time both converge: legacy files are adapted into the universal ComplianceFramework model (adapt_legacy_to_universal()), so the difference is about authoring cost and capabilities, not about what the rest of Prowler sees.

Side-by-side comparison

When to use which

Use universal when (any of these):

• The framework is new to Prowler — no existing attribute class, no existing formatter. This is the default: zero Python changes needed.
• The framework spans (or will span) more than one provider — DORA, CSA CCM, CIS Controls. One file covers all providers; extending to a new provider is a one-line checks edit.
• The attribute shape is unique to this framework — declare it in attributes_metadata instead of adding a Pydantic class to the Union.

Use legacy only when extending an existing legacy family:

• A new version of a shipped legacy framework (CIS 8.0 for AWS → new cis_8.0_aws.json, same CIS_Requirement_Attribute, same cis/ formatter).
• An existing legacy framework for a new provider (ENS for m365 → new ens_rd2022_m365.json + ens_m365.py transformer).
• Consistency with the family matters more than the universal benefits — a lone cis_8.0_aws in universal format while 20+ CIS files stay legacy would fragment the family.

Never: start a brand-new single-provider framework as legacy "because it's only AWS today". Universal handles single-provider fine (the checks dict just has one key) and you skip 3 output files + 3 registrations.

The same requirement in both schemas

Universal (prowler/compliance/my_framework_1.0.json):

{
  "framework": "My-Framework",
  "name": "My Framework 1.0",
  "version": "1.0",
  "description": "...",
  "attributes_metadata": [
    {"key": "Section", "type": "str", "required": true},
    {"key": "Service", "type": "str"}
  ],
  "outputs": {"table_config": {"group_by": "Section"}},
  "requirements": [
    {
      "id": "MF-1.1",
      "name": "Root MFA",
      "description": "Root account must have MFA enabled.",
      "attributes": {"Section": "IAM", "Service": "iam"},
      "checks": {
        "aws": ["iam_root_mfa_enabled"],
        "azure": []
      }
    }
  ]
}

Legacy (prowler/compliance/aws/my_framework_1.0_aws.json — plus a second file per extra provider, plus formatter + registrations):

{
  "Framework": "My-Framework",
  "Name": "My Framework 1.0 for AWS",
  "Version": "1.0",
  "Provider": "AWS",
  "Description": "...",
  "Requirements": [
    {
      "Id": "MF-1.1",
      "Name": "Root MFA",
      "Description": "Root account must have MFA enabled.",
      "Attributes": [
        {"ItemId": "MF-1.1", "Section": "IAM", "Service": "iam"}
      ],
      "Checks": ["iam_root_mfa_enabled"]
    }
  ]
}

Same control, but the universal file already covers Azure, validates its own attribute schema, and renders table/CSV/OCSF with no code. Field-by-field references for each schema follow below.

Architecture (Mental Model)

Prowler compliance is a four-layer system. Bugs usually happen where one layer doesn't match another, so know all four before touching anything.

Layer 1: SDK / Core Models — prowler/lib/check/

All in Pydantic v1 (from pydantic.v1 import ...). Three model groups live in compliance_models.py:

Legacy tree — Compliance → Compliance_Requirement / Mitre_Requirement:

• One *_Requirement_Attribute class per framework family. Registered today (Union order matters): ASDEssentialEight, CIS, ENS, ISO27001_2013, AWS_Well_Architected, KISA_ISMSP, Prowler_ThreatScore, CCC, C5Germany, CSA_CCM, STIG (Okta IDaaS), and Generic_Compliance_Requirement_Attribute as fallback.
• Generic MUST stay LAST in Compliance_Requirement.Attributes: list[Union[...]] — Pydantic v1 tries union members in order; Generic first would swallow every framework-specific attribute. NIST 800-53/CSF, PCI DSS, GDPR, HIPAA, SOC2, FedRAMP, SecNumCloud etc. intentionally use Generic.
• A root_validator rejects empty Framework, Provider or Name.
• MITRE uses the separate Mitre_Requirement model (Tactics, SubTechniques, Platforms, TechniqueURL at requirement top level, per-provider Mitre_Requirement_Attribute_{AWS,Azure,GCP}).

Universal tree — ComplianceFramework → UniversalComplianceRequirement:

• Flat attributes: dict per requirement, schema declared in attributes_metadata (key, label, type, enum, required, enum_display, enum_order, output_formats). A root_validator rejects missing required keys, unknown keys (drift guard), enum violations, and int/float/bool type mismatches. If attributes_metadata is omitted, no validation runs.
• checks: dict[provider, list[check_id]] — the provider list of the framework is derived from these keys (get_providers() / supports_provider()); the top-level provider field is only a fallback.
• outputs.table_config (group_by, split_by, scoring, labels) drives the CLI table; outputs.pdf_config exists in the model but is not consumed by the API PDF pipeline yet (see Layer 4).

Guardrails — Compliance_Requirement_ConfigConstraint:

• Fields Check, ConfigKey, Operator (lte|gte|eq|in|subset|superset), Value, optional Provider (required in universal multi-provider files).
• A root_validator rejects Value/Operator type mismatches at load time.
• Evaluation is centralized in prowler/lib/check/compliance_config_eval.py (evaluate_config_constraints, apply_config_status, get_effective_status, CONFIG_NOT_VALID_PREFIX = "Configuration not valid for this requirement."), shared by CSV/OCSF/table outputs and the API backend. A violated constraint forces the requirement to FAIL and prepends the reason to status_extended. Constraints whose ConfigKey is absent from audit_config are skipped (defaults assumed compliant).

Loaders:

• Compliance.get_bulk(provider) — legacy: scans only prowler/compliance/{provider}/ (+ external JSONs via the prowler.compliance entry-point group). Does NOT see top-level universal files.
• get_bulk_compliance_frameworks_universal(provider) — scans both the top-level prowler/compliance/ and every provider subdirectory, adapting legacy files via adapt_legacy_to_universal() (flattens Attributes[0] to a dict, wraps Checks as {provider: [...]}, infers attributes_metadata). Also loads external universal frameworks via the prowler.compliance.universal entry-point group (built-ins win collisions).
• get_check_compliance(finding, provider_type, bulk_checks_metadata) lives in prowler/lib/outputs/compliance/compliance_check.py (not in lib/check/compliance.py). It builds the per-finding dict keyed f"{Framework}-{Version}" only when Version is non-empty — an empty Version silently produces the key "{Framework}" and breaks downstream filters and tests.
• prowler/lib/check/compliance.py now contains only update_checks_metadata_with_compliance().

Layer 2: JSON Catalogs — prowler/compliance/

See "Compliance Catalog Coverage" below.

Layer 3: Output Formatters — prowler/lib/outputs/compliance/

Universal path (no Python needed per framework):

• universal/universal_table.py — get_universal_table(), renders the CLI table from outputs.table_config + attributes_metadata.
• universal/universal_output.py — UniversalComplianceOutput, builds the CSV Pydantic model dynamically from attributes_metadata.
• universal/ocsf_compliance.py — OCSFComplianceOutput; OCSF output is always generated for universal frameworks regardless of --output-formats.
• Orchestrated by process_universal_compliance_frameworks() in compliance.py, which runs before any legacy dispatch and removes the processed frameworks from the set.

Legacy path — per-framework directory, usually:

{framework}/
├── __init__.py
├── {framework}.py            # get_{framework}_table() summary-table function
├── {framework}_{provider}.py # One ComplianceOutput subclass per provider
└── models.py                 # One Pydantic CSV row model per provider

Directories today: asd_essential_eight, aws_well_architected, c5, ccc, cis, cisa_scuba, ens, generic, iso27001, kisa_ismsp, mitre_attack, okta_idaas_stig, prowler_threatscore, universal. Known deviations (don't "fix" them without a reason): iso27001/ has no table file (falls to the generic table), aws_well_architected/ has no per-provider files, cisa_scuba/ only ships googleworkspace.

• CSV writers emit ;-delimited files with UPPERCASE headers (ComplianceOutput.batch_write_data_to_file). Field names in models.py are public API — renaming breaks downstream consumers.
• Circular import rule: the table file ({framework}.py) must not import Finding directly or transitively (compliance.compliance → table module → ComplianceOutput → Finding → get_check_compliance → cycle). Keep table files bare (colorama, tabulate, prowler.config.config); when a module genuinely needs both, use if TYPE_CHECKING: or function-local imports (see universal_output.py / process_universal_compliance_frameworks).
• Legacy table functions have no docstrings; the universal ones do. Match the style of the file family you're touching.
• Dispatcher display_compliance_table() in compliance.py order: universal (table_config) first → cis_ → ens_ → mitre_attack → kisa → prowler_threatscore_ → c5_ → ccc_ → asd_essential_eight (substring) → okta_idaas_stig → else provider hook (provider.display_compliance_table(), may raise NotImplementedError) → get_generic_compliance_table(). iso27001, aws_well_architected and cisa_scuba ride the fallback on purpose.

Layer 4: API / UI

• API lazy loaders: api/src/backend/api/compliance.py — LazyComplianceTemplate / LazyChecksMapping (per-provider lazy caches over get_bulk_compliance_frameworks_universal, with Gunicorn background warm-up).
• API CSV export dispatch: COMPLIANCE_CLASS_MAP in api/src/backend/tasks/jobs/export.py, consumed from tasks/tasks.py. It is a dict provider → [(predicate, exporter_class)] with GenericCompliance as fallback. Predicates mix startswith for multi-version families (cis_, ens_, iso27001_, ccc_, cisa_scuba_, ...) and exact name == ... for true singletons (mitre_attack_aws, prowler_threatscore_*, asd_essential_eight_aws — and inconsistently c5_azure/c5_gcp, while aws uses startswith("c5_")). Rule of thumb: if the framework can ever grow versions or variants, use startswith.
• API overview ingestion: create_compliance_requirements() in api/src/backend/tasks/jobs/scan.py builds per-region rows from the lazy template and persists ComplianceRequirementOverview (COPY with bulk-create fallback) plus ComplianceOverviewSummary.
• API PDF reports: api/src/backend/tasks/jobs/reports/ — hardcoded FRAMEWORK_REGISTRY (own FrameworkConfig dataclass, NOT the SDK PDFConfig) with one generator class per framework. Only prowler_threatscore, ens, nis2, csa_ccm and cis have PDFs today; adding one means a generator class + registry entry + wiring in report.py.
• UI mapper routing: ui/lib/compliance/compliance-mapper.ts — getComplianceMappers() keyed by the JSON's framework value (e.g. "CIS", "CIS-Controls", "DORA", "Okta-IDaaS-STIG"). Unregistered frameworks fall back to the generic mapper + GenericCustomDetails automatically — a dedicated mapper/detail panel is a first-class upgrade, not a requirement to render.
• UI grouping varies per mapper: generic/cis group by Section/SubSection, iso by Category, ccc by FamilyName. All read attributes[0] — inconsistent values within one JSON become separate tree branches, so normalize before shipping.
• UI types: ui/types/compliance.ts — one *AttributesMetadata interface per framework, added to the AttributesItemData metadata union.
• UI icons: ui/components/icons/compliance/ + IconCompliance.tsx. Registration is an ordered substring match (COMPLIANCE_LOGOS): put framework-specific keywords before generic ones (nist before nis2, cisa before cis; aws deliberately last).

The CLI Pipeline (end-to-end)

prowler aws --compliance cis_7.0_aws          # framework key = JSON basename
    ↓
Compliance.get_bulk("aws")                      # legacy frameworks
get_bulk_compliance_frameworks_universal("aws") # legacy (adapted) + universal
    ↓
update_checks_metadata_with_compliance()        # attaches compliance to CheckMetadata
    ↓
execute_checks() → Finding objects
    ↓
get_check_compliance(finding, "aws", bulk)      # dict "{Framework}-{Version}" → [req_ids]
    ↓
process_universal_compliance_frameworks()       # universal: CSV + OCSF, then removed from set
per-provider elif branches in __main__.py       # legacy: AWSCIS(...).batch_write_data_to_file()
    ↓
display_compliance_table()                      # universal table first, then legacy elifs,
                                                # then generic fallback

Compliance Catalog Coverage

Counts as of 2026-07 (109 JSON files). Regenerate before trusting them:

for d in prowler/compliance/*/; do printf "%s: %s\n" "$(basename $d)" "$(ls $d*.json 2>/dev/null | wc -l)"; done
ls prowler/compliance/*.json   # universal, top-level

Universal (top-level, multi-provider): cis_controls_8.1.json (18 providers), csa_ccm_4.0.json (aws/azure/gcp/alibabacloud/oraclecloud), dora_2022_2554.json (aws/azure/gcp/alibabacloud/cloudflare).

Legacy per-provider (families, not exhaustive versions):

Providers with a compliance directory but no frameworks yet: cloudflare, iac, linode, llm, mongodbatlas, openstack, stackit. Provider keys inside universal checks dicts must match directory names under prowler/providers/ (lowercase).

Universal Schema Reference

Full spec in docs/developer-guide/security-compliance-framework.mdx. Skeleton:

{
  "framework": "DORA",
  "name": "Digital Operational Resilience Act (DORA) 2022/2554",
  "version": "2022/2554",
  "description": "Shown in --list-compliance and PDF reports.",
  "icon": "dora",
  "attributes_metadata": [
    {"key": "Pillar", "label": "Pillar", "type": "str", "required": true,
     "enum": ["ICT Risk Management", "..."],
     "output_formats": {"csv": true, "ocsf": true}},
    {"key": "Article", "type": "str", "required": true}
  ],
  "outputs": {
    "table_config": {"group_by": "Pillar"},
    "pdf_config": {"group_by_field": "Pillar", "charts": ["..."]}
  },
  "requirements": [
    {
      "id": "DORA-Art5",
      "name": "Governance and organisation",
      "description": "Requirement text verbatim from the source.",
      "attributes": {"Pillar": "ICT Risk Management", "Article": "Article 5"},
      "checks": {
        "aws": ["iam_no_root_access_key"],
        "azure": [],
        "gcp": []
      },
      "config_requirements": [
        {"Check": "iam_user_accesskey_unused", "Provider": "aws",
         "ConfigKey": "max_unused_access_keys_days", "Operator": "lte", "Value": 45}
      ]
    }
  ]
}

Universal fields, top level (ComplianceFramework)

Universal fields, per requirement (UniversalComplianceRequirement)

attributes_metadata entry fields (AttributeMetadata)

Key rules:

• --compliance key = JSON basename without .json (dora_2022_2554).
• Auto-discovered: no __init__.py, no formatter, no dispatcher registration.
• table_config.group_by, pdf_config.group_by_field and every charts[].group_by must reference a key declared in attributes_metadata.
• Runtime type validation only covers int/float/bool; str/list_str/ list_dict are documentation-only.
• Extending to a new provider = adding a key to requirement.checks. Nothing else.
• No automatic check-existence validation at load time — a typo'd check id silently produces a requirement with no findings. Always run the check-existence cross-check (see Validation).
• In universal files, always set Provider on every config constraint so a guardrail authored for an AWS check never affects Azure/GCP scans of the same requirement.

Legacy Schema Reference

Base legacy file structure:

{
  "Framework": "FRAMEWORK_NAME",
  "Name": "Full Framework Name with Version",
  "Version": "X.X",
  "Provider": "AWS",
  "Description": "Framework description...",
  "Requirements": [
    {
      "Id": "requirement_id",
      "Name": "Optional requirement name",
      "Description": "Requirement description",
      "Attributes": [ ... ],
      "Checks": ["check_name_1"],
      "ConfigRequirements": [ ... ]
    }
  ]
}

Legacy fields, top level (Compliance)

Legacy fields, per requirement (Compliance_Requirement)

MITRE files use Mitre_Requirement instead, which adds Tactics, SubTechniques, Platforms, TechniqueURL at the requirement top level.

Attribute shapes per framework family

Unlike universal (schema in-file), a legacy requirement's Attributes must match one of the Pydantic classes registered in Compliance_Requirement.Attributes — a shape matching no class silently falls through to Generic, dropping its specific fields. The most common shapes (full field sets in compliance_models.py):

CIS — cis_{version}_{provider}

{
  "Section": "1 Identity and Access Management",
  "SubSection": "Optional subsection",
  "Profile": "Level 1",
  "AssessmentStatus": "Automated",
  "Description": "...", "RationaleStatement": "...", "ImpactStatement": "...",
  "RemediationProcedure": "...", "AuditProcedure": "...",
  "AdditionalInformation": "...", "DefaultValue": "...", "References": "https://..."
}

Profile: Level 1|Level 2|E3 Level 1|E3 Level 2|E5 Level 1|E5 Level 2. AssessmentStatus: Automated|Manual.

ENS — ens_rd2022_{provider}

{
  "IdGrupoControl": "op.acc.1", "Marco": "operacional",
  "Categoria": "control de acceso", "DescripcionControl": "...",
  "Nivel": "alto", "Tipo": "requisito",
  "Dimensiones": ["trazabilidad", "autenticidad"],
  "ModoEjecucion": "automatico", "Dependencias": []
}

Nivel: opcional|bajo|medio|alto. Tipo: refuerzo|requisito|recomendacion|medida. Dimensiones: confidencialidad|integridad|trazabilidad|autenticidad|disponibilidad.

ISO 27001 — iso27001_{year}_{provider}

{
  "Category": "A.5 Organizational controls",
  "Objetive_ID": "A.5.1", "Objetive_Name": "Policies for information security",
  "Check_Summary": "Summary of what is being checked"
}

Note: Objetive_ID / Objetive_Name use this exact (mis)spelling.

MITRE ATT&CK — mitre_attack_{provider} (separate requirement model)

{
  "Name": "Exploit Public-Facing Application", "Id": "T1190",
  "Tactics": ["Initial Access"], "SubTechniques": [],
  "Platforms": ["IaaS"], "Description": "...",
  "TechniqueURL": "https://attack.mitre.org/techniques/T1190/",
  "Checks": ["guardduty_is_enabled"],
  "Attributes": [
    {"AWSService": "Amazon GuardDuty", "Category": "Detect",
     "Value": "Minimal", "Comment": "..."}
  ]
}

AzureService/GCPService for the other providers. Category: Detect|Protect|Respond. Value: Minimal|Partial|Significant.

CCC — ccc_{provider}

{
  "FamilyName": "Data", "FamilyDescription": "...",
  "Section": "CCC.Core.CN01 Encrypt Data for Transmission", "SubSection": "",
  "SubSectionObjective": "...",
  "Applicability": ["tlp-green", "tlp-amber", "tlp-red"],
  "Recommendation": "...",
  "SectionThreatMappings": [{"ReferenceId": "CCC", "Identifiers": ["CCC.Core.TH02"]}],
  "SectionGuidelineMappings": [{"ReferenceId": "NIST-CSF", "Identifiers": ["PR.DS-02"]}]
}

Applicability holds TLP tags (tlp-clear|tlp-green|tlp-amber|tlp-red).

ASD Essential Eight — asd_essential_eight_aws

{
  "Section": "Patch applications", "MaturityLevel": "ML1",
  "AssessmentStatus": "Automated", "CloudApplicability": "partial",
  "MitigatedThreats": ["..."], "Description": "...",
  "RationaleStatement": "...", "ImpactStatement": "...",
  "RemediationProcedure": "...", "AuditProcedure": "...",
  "AdditionalInformation": "...", "References": "..."
}

MaturityLevel: ML1|ML2|ML3. CloudApplicability: full|partial|limited|non-applicable.

DISA STIG — okta_idaas_stig_v1r2_okta

{
  "Section": "...", "Severity": "high", "RuleID": "...", "StigID": "...",
  "CCI": ["CCI-000015"], "CheckText": "...", "FixText": "..."
}

Severity: high|medium|low (maps to CAT I/II/III).

Other registered shapes

• AWS Well-Architected (aws_well_architected_framework_{pillar}_pillar_aws): Name, WellArchitectedQuestionId, WellArchitectedPracticeId, Section, SubSection, LevelOfRisk, AssessmentMethod, Description, ImplementationGuidanceUrl.
• KISA ISMS-P (kisa_isms_p_2023_{provider}): Domain, Subdomain, Section, AuditChecklist, RelatedRegulations, AuditEvidence, NonComplianceCases.
• C5 (c5_{provider}): Section, SubSection, Type, AboutCriteria, ComplementaryCriteria.
• CSA CCM (legacy shape; the shipped CSA CCM 4.0 is universal): Section, CCMLite, IaaS, PaaS, SaaS, ScopeApplicability.
• Prowler ThreatScore (prowler_threatscore_{provider}): Title, Section, SubSection, AttributeDescription, AdditionalInformation, LevelOfRisk (1–5), Weight (1/8/10/100/1000). Pillars: 1 IAM, 2 Attack Surface, 3 Logging and Monitoring, 4 Encryption. Available for aws, azure, gcp, kubernetes, m365, alibabacloud.
• Generic (fallback): ItemId, Section, SubSection, SubGroup, Service, Type, Comment — all optional. Used by NIST, PCI, GDPR, HIPAA, SOC2, FedRAMP, CISA, FFIEC, RBI, NIS2, GxP, SecNumCloud, etc.

Config Guardrails (ConfigRequirements)

Requirements backed by configurable checks can be silently "satisfied" by a loosened audit_config (e.g. CIS demands 45-day unused credentials but the scan ran with max_unused_access_keys_days: 120). Guardrails force such requirements to FAIL:

"ConfigRequirements": [
  {"Check": "iam_user_accesskey_unused",
   "ConfigKey": "max_unused_access_keys_days", "Operator": "lte", "Value": 45}
]

• Operators: lte/gte (numeric thresholds), eq (toggles/exact — use JSON booleans, not 0/1), in (scalar in allowed set), subset (allowlists — widening breaks it), superset (denylists — removing an entry breaks it).
• Value must be the strictest setting the control text tolerates.
• ConfigKey must be spelled exactly as the check reads it; unknown keys are silently skipped (defaults assumed OK).
• Guardrails only tighten (PASS→FAIL), never relax.
• Universal files: lowercase config_requirements + mandatory Provider per constraint.
• Tests: tests/lib/check/compliance_config_eval_test.py, compliance_config_constraint_model_test.py, compliance_config_requirements_data_test.py, plus per-output tests under tests/lib/outputs/compliance/.

Workflow A: Sync a Framework With an Upstream Catalog

Use when the framework is maintained upstream (CIS Benchmarks, FINOS CCC, CSA CCM, NIST, ENS, etc.) and Prowler needs to catch up.

Step 1 — Cache the upstream source

Download every upstream file to a local cache so iterations don't hit the network. For FINOS CCC:

mkdir -p /tmp/ccc_upstream
catalogs="core/ccc storage/object management/auditlog management/logging ..."
for p in $catalogs; do
  safe=$(echo "$p" | tr '/' '_')
  gh api "repos/finos/common-cloud-controls/contents/catalogs/$p/controls.yaml" \
    -H "Accept: application/vnd.github.raw" > "/tmp/ccc_upstream/${safe}.yaml"
done

Step 2 — Run the generic sync runner against a framework config

The sync tooling is three layers, so adding a framework only takes a YAML config (plus a parser module for an unfamiliar upstream format):

skills/prowler-compliance/assets/
├── sync_framework.py          # generic runner — works for any framework
├── configs/ccc.yaml           # per-framework config (canonical example)
└── parsers/finos_ccc.py       # parser module for FINOS CCC YAML

python skills/prowler-compliance/assets/sync_framework.py \
       skills/prowler-compliance/assets/configs/ccc.yaml

The runner loads the config, dynamically imports parser.module, calls parse_upstream(config) -> list[dict], then applies generic post-processing (id-uniqueness safety net, FamilyName normalization, legacy check-mapping preservation with config-driven fallback keys) and writes the provider JSONs with Pydantic post-validation.

To add a new framework sync:

1. Write assets/configs/{framework}.yaml (see ccc.yaml). Required sections:
   • framework — name, display_name, version (never empty — the runner refuses to start, because empty Version breaks the get_check_compliance() key), description_template.
   • providers — list of {key, display} pairs.
   • output.path_template — e.g. "prowler/compliance/{provider}/cis_{version}_{provider}.json".
   • upstream.dir — local cache (Step 1).
   • parser.module — module under parsers/; the rest of parser. is passed through opaque.
   • post_processing.check_preservation.primary_key (almost always Id) and fallback_keys — lists of Attributes[0] field names composed into tuples for recovering mappings when ids change. CCC: - [Section, Applicability]; CIS: - [Section, Profile]; NIST: - [ItemId]. List-valued fields are frozen to frozenset automatically.
   • post_processing.family_name_normalization (optional) — raw → canonical map; the UI groups by the exact attribute value, so upstream variants otherwise become separate tree branches.
2. Reuse an existing parser or write parsers/{name}.py implementing parse_upstream(config) -> list[dict] returning Prowler-format requirements with guaranteed-unique ids. The runner raises on duplicates — it never silently renumbers, because mutating a canonical upstream id (CIS 1.1.1, NIST AC-2(1)) would be catastrophic. The parser owns all upstream quirks: foreign-prefix rewriting, genuine collision renumbering, multi-shape handling.

Gotchas the runner already handles (from the FINOS CCC v2025.10 sync):

• Multiple upstream YAML shapes. Most FINOS CCC catalogs use control-families: [...] but storage/object uses top-level controls: [...]. A single-shape parser silently drops entire catalogs — this exact bug dropped ObjStor for a full iteration. Test with one file of each shape.
• Whitespace collapse. Upstream | block scalars keep newlines; Prowler stores single-line. Collapse with " ".join(value.split()).
• Foreign-prefix id rewriting. Upstream aliases requirements across catalogs keeping the original prefix (CCC.AuditLog.CN08.AR01 nested under CCC.Logging.CN03) — rewrite to fit the parent (CCC.Logging.CN03.AR01).
• Genuine upstream collisions. Two different requirements sharing one id (upstream typo): renumber the second to the next free number; check-mapping preservation recovers by the fallback keys.
• Populate Version — fail-fast beats the silent broken-key bug.

Step 3 — Validate before committing

Run the full Validation section below (universal loader + check existence + CLI smoke + pytest).

Step 4 — Add an attribute model if needed

Only if the framework has fields beyond Generic_Compliance_Requirement_Attribute and must stay legacy. Add the class to compliance_models.py and register it in the Compliance_Requirement.Attributes Union before Generic (Generic stays last). For new frameworks, prefer universal attributes_metadata instead.

Workflow B: Audit Check Mappings as a Cloud Auditor

Use when the user asks to review existing mappings. This is the highest-value compliance task — it surfaces padded mappings with zero actual coverage and missing mappings for legitimate coverage.

The golden rule

A Prowler check's title/risk MUST literally describe what the requirement text says. "Related" is not enough. If no check actually addresses the requirement, leave the checks list empty (MANUAL) — honest MANUAL is worth more than padded coverage.

Audit process

1. Build a per-provider check inventory — assets/build_inventory.py (writes /tmp/checks_{provider}.json for every provider discovered under prowler/providers/).

2. Query it — assets/query_checks.py (run from the repository root):

   python skills/prowler-compliance/assets/query_checks.py aws encryption transit  # keyword AND-search
   python skills/prowler-compliance/assets/query_checks.py aws --service iam       # all iam checks
   python skills/prowler-compliance/assets/query_checks.py aws --id kms_cmk_rotation_enabled

3. Dump a framework section with current mappings — assets/dump_section.py:

   python skills/prowler-compliance/assets/dump_section.py ccc "CCC.Core."
   python skills/prowler-compliance/assets/dump_section.py cis_5.0_aws "1."

4. Encode explicit REPLACE decisions — assets/audit_framework_template.py:

   DECISIONS = {}
   DECISIONS["CCC.Core.CN01.AR01"] = {
       "aws": ["cloudfront_distributions_https_enabled", ...],
       "azure": ["storage_secure_transfer_required_is_enabled", ...],
       "gcp": ["cloudsql_instance_ssl_connections"],
       # Missing provider key = leave the legacy mapping untouched
   }
   # Empty list = EXPLICITLY MANUAL (overwrites legacy)
   DECISIONS["CCC.Core.CN01.AR07"] = {"aws": [], "azure": [], "gcp": []}

   REPLACE, not PATCH. Full lists make the audit reproducible and surface hidden assumptions in the legacy data.

5. Pre-validate every check id against the inventory; the script MUST abort with stderr listing typos (real audits caught storage_secure_transfer_required_enabled → storage_secure_transfer_required_is_enabled, sqlserver_minimum_tls_version_12 → sqlserver_recommended_minimal_tls_version, and several checks that simply don't exist).

6. Apply + validate + test:

   python /path/to/audit_script.py
   uv run pytest -n auto tests/lib/outputs/compliance/ tests/lib/check/ -q

For the curated mapping table (requirement text → AWS/Azure/GCP checks) and the list of controls Prowler genuinely cannot verify, see references/check-mapping-reference.md.

Workflow C: Add a New Universal Framework

1. Author prowler/compliance/{framework}_{version}.json following the Universal Schema Reference above (use dora_2022_2554.json or csa_ccm_4.0.json as template).
2. Declare every attribute in attributes_metadata (with required/enum where possible — that's your load-time validation) and a outputs.table_config.group_by.
3. Map checks per provider; add config_requirements (with Provider) for configurable checks; leave empty lists for manual requirements — include every requirement of the source catalog (coverage percentages depend on the full denominator).
4. Validate (section below). No Python registration of any kind is needed for CLI table/CSV/OCSF.
5. Optional first-class UI: mapper in ui/lib/compliance/{framework}.tsx, registration in getComplianceMappers() under the JSON's framework value, detail panel, *AttributesMetadata type, and icon (ordered keyword!). Until then the generic mapper renders it.
6. Optional API extras: CSV exporter entry in COMPLIANCE_CLASS_MAP; PDF generator + FRAMEWORK_REGISTRY entry if a PDF is required.
7. Tests: extend tests/lib/check/universal_compliance_models_test.py with a case loading the new JSON. The parametrized test_loads_as_universal already picks the file up automatically.
8. Changelog fragment prowler/changelog.d/<slug>.added.md + user-guide tutorial under docs/user-guide/compliance/tutorials/ for high-profile frameworks.

Workflow D: Add a New Legacy Output Formatter

Only for new members of an existing legacy family. Follow the c5/ or ccc/ layout exactly:

1. mkdir prowler/lib/outputs/compliance/{framework} with __init__.py.
2. {framework}.py — copy c5/c5.py, change function name + framework string; the diff should be just those lines. No docstring (legacy style).
3. models.py — one Pydantic CSV row model per provider. Column sets differ per provider (AccountId/Region vs SubscriptionId/Location vs ProjectId/Location); per-provider files are the convention — don't collapse them into a parameterized class, reviewers will reject it.
4. {framework}_{provider}.py — {Framework}_{Provider}(ComplianceOutput) with transform(); this file may import Finding.
5. Register:
   • compliance.py → display_compliance_table() elif branch (+ top import).
   • prowler/__main__.py → per-provider elif compliance_name.startswith(...) branches instantiating the writer classes.
   • api/src/backend/tasks/jobs/export.py → COMPLIANCE_CLASS_MAP entries (startswith for families, exact match only for true singletons).
6. Tests under tests/lib/outputs/compliance/{framework}/ + fixtures in tests/lib/outputs/compliance/fixtures.py (1 evaluated + 1 manual requirement to exercise both transform() paths).

Circular import warning: the table file must not import Finding directly or transitively (cycle: compliance.compliance → table → ComplianceOutput → Finding → get_check_compliance → compliance.compliance). Keep it bare; use TYPE_CHECKING/function-local imports where both are genuinely needed.

Validation (run before every commit)

1. Schema load (both formats):

   from prowler.lib.check.compliance_models import (
       load_compliance_framework_universal,
       get_bulk_compliance_frameworks_universal,
   )
   fw = load_compliance_framework_universal("prowler/compliance/<file>.json")
   assert fw is not None, "check logs for the ValidationError"
   print(fw.framework, len(fw.requirements), fw.get_providers())
   assert "<file_basename>" in get_bulk_compliance_frameworks_universal("aws")

   Remember: the universal loader is lenient (skips broken files with a log line) — an assert fw is not None is mandatory, a green scan is not proof.

2. Check existence — no loader validates this; a stale id is silent dead weight:

   import json
   from pathlib import Path
   for prov in ["aws", "azure", "gcp"]:
       real = {p.stem.replace(".metadata", "")
               for p in Path(f"prowler/providers/{prov}/services").rglob("*.metadata.json")}
       data = json.load(open(f"prowler/compliance/{prov}/<file>.json"))
       refs = {c for r in data["Requirements"] for c in r["Checks"]}
       missing = refs - real
       assert not missing, f"{prov} missing: {missing}"

   (For universal files use r.get("checks", {}).get(prov, []) instead — requirements may legitimately omit a provider key.)

3. CLI smoke test:

   uv run python prowler-cli.py <provider> --list-compliance          # appears?
   uv run python prowler-cli.py <provider> --compliance <key> --log-level ERROR

   Verify the CSV under output/compliance/, the summary table sections, and the findings roll-up.

4. Tests:

   uv run pytest -n auto tests/lib/check/universal_compliance_models_test.py \
     tests/lib/outputs/compliance/

   test_loads_as_universal is parametrized over every JSON in prowler/compliance/ (top-level + subdirectories) — a malformed file fails CI here even if you never wrote a dedicated test.

5. What CI/pre-commit do and don't cover: pre-commit only guarantees well-formed/pretty JSON (check-json, pretty-format-json) — no semantic validation. The workflow .github/workflows/pr-check-compliance-mapping.yml flags PRs adding new checks without mapping them to any framework (label needs-compliance-review; skip with label no-compliance-check). Semantic validation happens in the pytest suite above and manually via skills/prowler-compliance-review/assets/validate_compliance.py (note: that validator assumes the legacy schema).

6. Prowler Local Server: docker compose up and confirm the compliance page renders requirements, sections and widgets.

Conventions and Hard-Won Gotchas

1. Universal first. A new framework that starts as legacy needs 3 output files + 3 registrations; the same framework as universal needs zero. Only extend legacy families.
2. Generic_Compliance_Requirement_Attribute stays LAST in the legacy Attributes Union — Pydantic v1 tries members in order; Generic first silently swallows every specific shape.
3. Pydantic v1 everywhere in compliance_models.py (from pydantic.v1 import ...). Don't mix in v2.
4. get_check_compliance() lives in prowler/lib/outputs/compliance/compliance_check.py and keys the dict f"{Framework}-{Version}" only when Version is non-empty. Never ship Version: "" — the key silently degrades to "{Framework}" and breaks filters, tests and --compliance. For legacy files the filename version substring must match Version (the CLI reads compliance_framework.split("_")[1]).
5. Compliance.get_bulk() does not see top-level universal files — only get_bulk_compliance_frameworks_universal() does. Wire new code paths against the universal loader.
6. Loader leniency differs: legacy loader exits the process on a broken JSON; universal loader logs and skips. A missing framework after your edit usually means the universal loader dropped it — check the logs.
7. Circular import protection: legacy table dispatcher files must not import Finding (directly or transitively). Use TYPE_CHECKING or function-local imports when a module needs both sides (that's how the universal formatter does it).
8. Per-provider formatter files are the legacy convention — but know the exceptions before flagging them (iso27001 has no table file, aws_well_architected has no per-provider files, cisa_scuba is googleworkspace-only). CSV model field names are public API.
9. CSV output: ; delimiter, UPPERCASE headers. OCSF compliance output is always generated for universal frameworks regardless of --output-formats.
10. COMPLIANCE_CLASS_MAP mixes predicate styles: startswith for multi-version families, exact == for singletons. When in doubt use startswith — exact match blocked versioned CCC variants until 2026.
11. UI grouping is per-mapper, always on attributes[0]: generic/cis → Section/SubSection, iso → Category, ccc → FamilyName. Inconsistent values (or empty Section) create orphan/duplicate tree branches — normalize before shipping.
12. UI has a generic fallback — an unregistered framework still renders. A dedicated mapper/panel/icon is an upgrade, not a prerequisite.
13. Icon registration is ordered substring matching in IconCompliance.tsx — specific keywords before generic (nist before nis2, cisa before cis, aws last).
14. API PDF pipeline is not PDFConfig-driven yet — it has its own FRAMEWORK_REGISTRY (5 frameworks). Don't assume adding pdf_config to a JSON produces a PDF in Prowler App.
15. Pre-validate every check id against the per-provider inventory before writing JSON. No loader will catch a typo; the requirement just never matches a finding.
16. REPLACE beats PATCH for audit decisions — full explicit lists are reproducible and surface legacy assumptions.
17. When no check applies, MANUAL is correct. Don't pad mappings with tangential checks; compliance reports must stay actionable.
18. Include every requirement of the source catalog, automated or not — compliance percentages use the full requirement count as denominator.
19. Provider coverage is asymmetric (AWS dense; Azure/GCP thinner; new providers minimal). Accept it — don't force parity Prowler can't verify.
20. Guardrail authoring: strictest tolerated Value, exact ConfigKey spelling, Provider mandatory in universal files, booleans as JSON booleans. Malformed constraints are treated as satisfied — validate with the config tests, don't trust silence.

Useful One-Liners

# Find duplicate requirement IDs (legacy | universal)
jq -r '.Requirements[].Id' file.json | sort | uniq -d
jq -r '.requirements[].id' file.json | sort | uniq -d

# Count manual requirements (legacy | universal, per provider)
jq '[.Requirements[] | select((.Checks | length) == 0)] | length' file.json
jq '[.requirements[] | select((.checks.aws // [] | length) == 0)] | length' file.json

# List unique check references (legacy | universal)
jq -r '.Requirements[].Checks[]' file.json | sort -u
jq -r '.requirements[].checks[]? | .[]' file.json | sort -u

# Providers covered by a universal framework
jq '[.requirements[].checks | keys[]] | unique' file.json

# Spot inconsistent grouping values (UI tree branches)
jq '[.Requirements[].Attributes[0].Section] | unique' file.json
jq '[.Requirements[].Attributes[0].FamilyName] | unique' file.json

# Requirements with config guardrails (empty arrays are truthy in jq — check length)
jq '[.Requirements[] | select((.ConfigRequirements // []) | length > 0)] | length' file.json

# Diff requirement ids between two versions
diff <(jq -r '.Requirements[].Id' a.json | sort) <(jq -r '.Requirements[].Id' b.json | sort)

# Where is a check mapped across all frameworks?
grep -rl "my_check_name" prowler/compliance/

# Does a check exist?
find prowler/providers/aws/services -name "{check_id}.metadata.json"

# Validate one file with the universal loader
python -c "from prowler.lib.check.compliance_models import load_compliance_framework_universal as l; fw=l('prowler/compliance/aws/cis_7.0_aws.json'); print(fw.framework, len(fw.requirements))"

Commands

prowler {provider} --list-compliance
prowler {provider} --compliance cis_7.0_aws
prowler aws --compliance cis_7.0_aws pci_4.0_aws
prowler aws --compliance dora_2022_2554            # universal key = file basename
prowler aws --list-compliance-requirements cis_7.0_aws
prowler aws --compliance cis_7.0_aws -M csv json html

Code References

Layer 1 — SDK / Core

• prowler/lib/check/compliance_models.py — legacy + universal model trees, Compliance_Requirement_ConfigConstraint, all loaders and the legacy→universal adapter
• prowler/lib/check/compliance.py — update_checks_metadata_with_compliance
• prowler/lib/check/compliance_config_eval.py — guardrail evaluation (shared with the API)
• prowler/lib/outputs/compliance/compliance_check.py — get_check_compliance
• prowler/lib/check/utils.py — list_compliance_modules

Layer 2 — JSON Catalogs

• prowler/compliance/*.json — universal, multi-provider (auto-discovered)
• prowler/compliance/{provider}/ — legacy, per-provider (auto-discovered)

Layer 3 — Output Formatters

• prowler/lib/outputs/compliance/universal/ — universal_table.py, universal_output.py, ocsf_compliance.py
• prowler/lib/outputs/compliance/{framework}/ — legacy per-framework packages
• prowler/lib/outputs/compliance/compliance.py — process_universal_compliance_frameworks, display_compliance_table
• prowler/lib/outputs/compliance/compliance_output.py — ComplianceOutput base + CSV writer
• prowler/__main__.py — universal processing + per-provider legacy writer branches

Layer 4 — API / UI

• api/src/backend/api/compliance.py — LazyComplianceTemplate, LazyChecksMapping, cache warm-up
• api/src/backend/tasks/jobs/export.py — COMPLIANCE_CLASS_MAP
• api/src/backend/tasks/jobs/scan.py — create_compliance_requirements (overview ingestion)
• api/src/backend/tasks/jobs/reports/ — PDF generators + FRAMEWORK_REGISTRY
• ui/lib/compliance/compliance-mapper.ts — mapper routing + generic fallback
• ui/lib/compliance/{framework}.tsx — per-framework mappers
• ui/components/compliance/compliance-custom-details/ — detail panels
• ui/types/compliance.ts — attribute metadata types
• ui/components/icons/compliance/ + IconCompliance.tsx — icons (ordered)

Tests

• tests/lib/check/universal_compliance_models_test.py — includes the parametrized test_loads_as_universal over every shipped JSON
• tests/lib/check/compliance_check_test.py, compliance_config_eval_test.py, compliance_config_constraint_model_test.py, compliance_config_requirements_data_test.py, mitre_config_requirements_test.py
• tests/lib/outputs/compliance/ — per-framework + universal + dispatcher + config-status coverage tests; shared fixtures.py

Resources

• Docs (source of truth for contributors): docs/developer-guide/security-compliance-framework.mdx (both schemas, guardrails, validation, PR process), docs/user-guide/compliance/tutorials/compliance.mdx, docs/user-guide/compliance/tutorials/cross-provider-compliance.mdx
• Repo tooling (util/compliance/): CSV→JSON generators (generate_json_from_csv/), ccc/from_yaml_to_json.py, compliance_mapper/, threatscore/
• Skill assets (assets/):
  • sync_framework.py + configs/ccc.yaml + parsers/finos_ccc.py — config-driven upstream sync (Workflow A)
  • build_inventory.py, query_checks.py, dump_section.py, audit_framework_template.py — audit tooling (Workflow B)
  • Legacy JSON templates: cis_framework.json, ens_framework.json, iso27001_framework.json, mitre_attack_framework.json, prowler_threatscore_framework.json, generic_framework.json
• References: references/compliance-docs.md — model/loader quick reference; references/check-mapping-reference.md — curated requirement-text → checks mapping table + honest-MANUAL list
• Sister skill: prowler-compliance-review — PR review checklist + validate_compliance.py (legacy-schema validator)
• After editing this skill's frontmatter, run ./skills/skill-sync/assets/sync.sh to regenerate the AGENTS.md auto-invoke tables.