AI Governance
Self-assessed maturity of AEGIS governance templates and controls against NIST AI RMF, the EU AI Act, ISO 42001, SOC 2, and FedRAMP High. No external audit has been conducted.
This page is a self-assessment of AEGIS governance templates and controls against several frameworks. No external audit, certification, or review has been conducted on AEGIS.
Framework Maturity Assessment
| Maturity Level | Definition |
|---|---|
| Nascent | Awareness exists; no artifacts or controls implemented |
| Designed | Artifacts created, procedures documented, not yet operationalized |
| Operational | Controls enforced in CI/CD, evidence collected, procedures executed |
| Validated | External audit or assessment confirms effectiveness |
Current Maturity by Framework
| Framework | Maturity | What's Enforced by CI | What's Documented Only | Key Gaps |
|---|---|---|---|---|
| NIST AI RMF 1.0 | Designed → Operational | Risk register validated, model card schema checked, artifacts present in CI | MAP/MEASURE functions partially covered | MEASURE subcategories need deeper metrics |
| EU AI Act | Designed | Technical file structure checked by CI (tools/ci/ai_act_lint.py); TEV report self-authored, not externally validated | Conformity assessment not started | No notified body engagement; TEV report needs external validation |
| ISO 42001 | Designed | AIMS policy documented, PDCA cycle framework | Only Plan and Check phases have substance | Do and Act phases need operational evidence |
| SOC 2 | Designed | Quality gates enforce CC8.1 (change management); RBAC enforces CC6.1 | Runbook templates for trust criteria | No SOC 2 audit conducted; runbooks not yet executed |
| FedRAMP High | Nascent → Designed | BCP/DRP and IRP documented | Access review procedures, vendor risk assessment | FedRAMP High is the target baseline. Formal 3PAO process not yet started; technological alignment in progress. FIPS 140-3 validated PQC modules unavailable industry-wide (earliest Q2 2027) |
What's Real and Strong
These capabilities are implemented, tested, and enforced:
- Cryptography: BIP-322 dual signatures, Ed25519, ML-DSA-65 (post-quantum), ML-KEM-768, HybridKEM — all implemented with 500+ tests
- RBAC: Fail-closed RBAC enforcement with 8 constraint types and a four-eyes principle
- PII encryption: 12 fields encrypted at rest using HybridKEM (X25519 + ML-KEM-768), 850+ lines of implementation
- Two-key overrides: Enforced via dual signature verification; each signature is bound to its role and signer id (P5-1). Caveat: signer public keys are still supplied by the caller rather than enrolled per signer (ROADMAP G109), so distinct custody is not yet cryptographically enforced
- Quality gates: Blocking in CI — ruff, mypy, bandit, pytest with 90% coverage floor
- Hash-chained audit trail: SHA-256 hash-chained telemetry events; chain verification detects edited entries. Tail-deletion detection applies to the persisted decision chain (
AuditChainStore), not to telemetry events - Audit trail: Comprehensive logging coverage for all evaluation paths with structured telemetry pipeline
Operational Maturity
Operational procedures (access reviews, BCP/DRP drills, incident response exercises) are scheduled for Q2 2026. Contact the AEGIS team for current status and the compliance remediation tasks (C3-C5).
Standards Alignment
| Standard | Coverage | Key Controls |
|---|---|---|
| NIST AI RMF 1.0 | GOVERN, MAP, MEASURE, MANAGE | Risk register, model card, data card, oversight plan. 45/72 subcategories fully implemented across all 4 functions (63% full, 86% including partial) |
| EU AI Act | Art. 10, 11, 14, 16, 72 | Self-authored system register, technical file template, human-oversight plan, and post-market monitoring plan |
| ISO 42001 | AIMS Policy (PDCA) | AI management system policy and continuous improvement |
| OWASP Agentic Top 10 | ASI01-ASI10 | Full control matrix with verification methods |
| CoSAI MCP Threat Model | MCP-T1 through MCP-T12 | 9/12 STRONG, 2/12 MODERATE, 1/12 PARTIAL |
Governance Artifacts
AEGIS maintains these governance artifacts in the ai/ directory:
| Artifact | Purpose | Standard | CI Validated |
|---|---|---|---|
system-register.yaml | AI system inventory | EU AI Act Art. 16 | Yes (schema + required fields) |
risk-register.yaml | Risk identification and mitigation | ISO 31000 / NIST AI RMF | Yes (schema + controls present) |
model-card.yaml | Decision engine characteristics | EU AI Act Art. 11 | Yes (schema + required fields) |
data-card.yaml | Telemetry data governance | EU AI Act Art. 10 | Yes (schema + required fields) |
oversight-plan.md | Human oversight procedures | EU AI Act Art. 14 | Yes (content validation) |
postmarket-monitoring.md | Continuous monitoring plan | EU AI Act Art. 72 | Yes (content validation) |
AIMS-POLICY.md | AI management system policy | ISO 42001 | Yes (content validation) |
technical_file/testing/tev_report.md | Test evidence report (self-authored) | EU AI Act | Yes (non-template check) |
SOC 2 / FedRAMP Inheritance
AEGIS provides governance controls that support downstream compliance:
- Audit trails: Hash-chained decision logs with SHA-256 chain verification that detects tail deletion and entry edits (CC7.2, CC7.3); not independently anchored
- Access control: RBAC with fail-closed enforcement (CC6.1, CC6.3)
- Change management: Quality gates, CI/CD pipeline, approval workflows (CC8.1)
- Risk assessment: Quantitative risk gates with Bayesian confidence (CC3.2)
- Monitoring: Prometheus metrics, alerting, drift detection (CC7.1)
- Cryptography: Post-quantum resistant signatures and encryption (CC6.7)
Organizations using AEGIS can reference these controls in their own SOC 2 or FedRAMP documentation as inherited controls from the governance layer.
Current limitation: No SOC 2 Type 2 or FedRAMP audit has been conducted on AEGIS itself. These controls are designed for inheritance, not as evidence of AEGIS's own certification. FedRAMP High is the target baseline; formal authorization process has not started. A FIPS-approved classical crypto path (ECDSA P-256 / ECDH P-384) is planned for FedRAMP deployments alongside the existing post-quantum hybrid. Note: as of Q1 2026, no FIPS 140-3 validated PQC module exists industry-wide (earliest projected: Q2 2027).
Operational Runbooks
Compliance runbooks are maintained in docs/compliance/:
| Runbook | Purpose | Status |
|---|---|---|
| Business Continuity / DR | RPO/RTO targets and failover procedures | Documented; first drill pending Q2 2026 |
| Incident Response | Security incident handling | Documented; tabletop exercise pending |
| Access Review | Periodic access certification | Documented; first quarterly review pending Q2 2026 |
| Vendor Risk | Third-party risk management (AWS) | Completed for AWS; other vendors pending |
| Change Management | Change approval and rollback | Operational (enforced via CI quality gates) |
| Data Subject Requests | GDPR/privacy request handling | Documented; no requests received (internal-only deployment) |