AegisSovereignAI

AegisSovereignAI: Trusted AI for the Distributed Enterprise

Executive Summary

In a Distributed Enterprise, Infrastructure Security (Layer 1 in Figure 1) and AI Governance (Layer 3 in Figure 1) are often loosely coupled across the spectrum from centralized clouds to the far edge. This fragmentation results in a dangerous "Accountability Gap" where workload/user identities are easily spoofed, compliance creates massive Personally Identifiable Information (PII) liability, and compromised infrastructure — whether in a hyperscale data center or a remote branch office — can feed fake data to applications undetected.

AegisSovereignAI bridges this gap by serving as a unifying control plane for the comprehensive distributed footprint. Through a Unified and Extensible Identity (Layer 2 in Figure 1) framework, it cryptographically fuses workloads/user identities using silicon-level attestation with application-level governance while preserving privacy to create a single, cohesive identity architecture that extends from the Cloud Core to the Far Edge.

This transforms AI security from "Best-Effort" Zero-trust to Privacy-First Verifiable Intelligence — enabling cryptographic proof of compliance (data residency, prompt governance, output filtering) without disclosing sensitive PII or proprietary logic. This ensures that sensitive data (financial, medical, etc.) is processed only when the hardware, the location, and the workload/user identity are simultaneously verified, providing end-to-end sovereignty across the entire enterprise estate.

Figure 1: AegisSovereignAI Architecture Summary - Bridging Infrastructure, Identity, and Governance.

See the [Unified Identity Hybrid Cloud PoC](./hybrid-cloud-poc/README.md) — run ./run-demo.sh to see the full trust chain in action (~7 min).

Enterprise Sovereign Use Cases (Focus: High-Security/Compliance Sectors e.g., Banking, Healthcare, Defense/Government)

1. The Enterprise Customer (High-Security/Compliance End-Consumer e.g., High-Net-Worth Client)

• Core Use Case: Private Wealth Gen-AI Advisory (Unmanaged Devices). Providing high-net-worth clients with AI-driven portfolio insights on their personal, unmanaged devices while using their physical location for Regulation K (Reg-K) compliance without disclosing precise location to the AI service.

2. The Enterprise Employee (Regulated Sector Employee e.g., Branch Relationship Manager)

• Core Use Case: Secure Remote Branch Operations. Allowing Relationship Managers to access sensitive PII from "Green Zone" servers on managed hardware, whether at a branch or a verified remote location.

3. The Enterprise Tenant (Line-of-Business Owner aka LOB e.g., Mortgage and Credit Card)

• Core Use Case: Secure Sandboxing for LOBs. Enabling enterprise tenants (e.g., Mortgage and Credit Card) to share the same physical Sovereign Cloud while ensuring total cryptographic isolation of their respective workloads, including AI models and data. From a tenant AI service perspective: data ingestion pipelines must prove PII was redacted and provenance verified before entering the tenant's vector store; the AI system prompt must contain mandatory safety guardrails (e.g., "never disclose account numbers"); user prompts must be scanned for injection attacks (e.g. "ignore previous instructions"); and AI outputs must be verified for PII leakage (hallucinations) before delivery.

4. The Regulator (e.g., Office of the Comptroller of the Currency (OCC), European Central Bank (ECB), or Securities and Exchange Commission (SEC))

• Core Use Case: Automated Regulatory Audit. While traditional audit models provide visibility through coarse data logging, applying this to AI creates a Privacy Liability Paradox: the more granular the audit (e.g., logging raw prompts/outputs), the higher the ingestion risk of sensitive PII and proprietary secrets. The Regulator requires real-time, cryptographically verifiable proof-of-compliance—demonstrating that: (1) all data ingested into AI systems (training data, Retrieval-Augmented Generation / RAG vector stores) was properly redacted and provenance-verified; (2) every AI interaction across the Enterprise strictly followed mandatory policy (trusted hardware, untampered models, and data residency); all without the liability of raw data ingestion or the exposure of proprietary prompt logic. This supports the reproducibility and documentation principles required by the Model Risk Management (MRM) regulatory framework and Federal Reserve/OCC Supervisory Letter SR 11-7 (Interagency Guidance on Model Risk Management).

Technical Challenges for Addressing Use Cases

To address the above use cases, we must solve the unique technical problems below. Note that the below technical problems are not unique to AI or Financial Services but are especially critical for the security, privacy, and compliance of the above use cases.

1. The Fragility of Identity & Geofencing

Traditional security relies on bearer tokens and IP-based geofencing, which are fundamentally non-binding and easily spoofed.

• Replay Attacks: Standard tokens function like a physical key; if a malicious actor intercepts a token, they can replay it to impersonate a legitimate workload (e.g., an AI agent).
• VPN-based Spoofing: Commonly used IP-based location checks are trivial to bypass using VPNs, allowing remote attackers to appear within "Green Zones."
• Example (Use Case 2 - Enterprise Employee): A Relationship Manager attempts to access the "Green Zone" server from an unauthorized jurisdiction via a residential VPN. Traditional IP-checks fail to detect the spoofed location.

2. The Residency vs. Privacy Deadlock

Regulators require proof of data residency (e.g., Regulation K aka Reg-K), but traditional geofencing relies on ingesting high-resolution location data (GPS, Mobile Network, etc.), creating massive PII liability under privacy regulations (e.g., General Data Protection Regulation (GDPR)). Enterprises are often forced to choose between non-compliance or privacy violation.

• Example (Use Case 1 - Enterprise Customer): A high-net-worth client uses the Private Wealth Gen-AI Advisory from their personal mobile device. The organization (e.g., bank) must prove to an EU regulator that the AI inference stayed within the EEA (Reg-K compliance), but doing so requires ingesting or storing raw GPS data from the client's device — a GDPR violation.

3. Infrastructure Compromise

Modern AI workloads are vulnerable to infrastructure compromise, where a compromised OS or Hypervisor feeds, for example, fake location data to the application (e.g., via Frida hooks), tricking compliance logic while the device is in an unauthorized jurisdiction.

• Example (Use Case 2 - Enterprise Employee): A compromised branch server's hypervisor feeds fake "within Green Zone" location data to the AI workload via Frida hooks, allowing a Relationship Manager to appear compliant while accessing sensitive PII from an unauthorized jurisdiction.

4. The "Silicon Lottery": Hardware-Induced Drift & Computational Determinism

AI prompt response drift can be influenced by the type of hardware. Response can vary based on randomness setting (e.g., temperature). Even when randomness is fully disabled (e.g., temperature=0), the same model can produce different outputs on different hardware types (e.g., NVIDIA A100 vs H100) due to floating-point math and parallel execution differences. For quantitative risk management, Computational Determinism — ensuring that the same model on the same hardware type produces consistent results — is essential. Enterprises require the ability to restrict and verify hardware types to ensure deterministic outcomes for regulated workloads.

• Example (Use Case 3 - Enterprise Tenant): The Mortgage LOB's credit risk model produces different risk scores when run on A100 vs H100 GPUs due to floating-point variations. Traditional infrastructure management cannot guarantee which hardware type executed a given inference, making regulatory reproducibility within the organization impossible.

5. The Black-Box Governance Gap: Integrity & Data Liability

AI models are non-deterministic, making them difficult to audit. There is no cryptographic proof that a specific decision was made using untampered AI models/prompts without disclosing sensitive data. This is further complicated by Prompt Injection (malicious instructions) and Hallucinations (unintended PII leakage).

• The "Audit Paradox": Traditional logging for compliance creates massive PII/IP liability, but not logging prevents forensics and "Effective Challenge."
• Example (Use Case 3 & 4 - Enterprise Tenant & Regulator): An OCC auditor needs to verify that the Credit Card LOB's AI agent didn't use prohibited demographic data for credit scoring. Under current methods, the organization must disclose raw prompts to the auditor — revealing the LOB's proprietary scoring logic and customer PII — creating significant liability.

6. Bring Your Own Device (BYOD) Security Gaps

BYOD devices are unmanaged and unverified, making them a significant security risk for data leakage and unauthorized access.

• Example (Use Case 1 - Enterprise Customer): A high-net-worth client accesses the Private Wealth Gen-AI Advisory from their personal iPad. The device may be jailbroken or compromised without the organization's knowledge, creating an undetectable data leakage vector for sensitive portfolio information.

7. Edge Security Gaps

Edge nodes are often in untrusted physical locations, making them vulnerable to physical tampering and unauthorized environment modification.

• Example (Use Case 2 - Enterprise Employee): A branch server used by Relationship Managers is physically compromised or stolen. Traditional software-based security cannot detect hardware tampering, allowing attackers to extract AI model weights and sensitive customer PII.

The Three-Layer Trust Architecture: Fusing Silicon, Identity, and Governance

AegisSovereignAI bridges Infrastructure Security (Layer 1 in Figure 2) and AI Governance (Layer 3 in Figure 2) by serving as a unifying control plane. Through a Unified and Extensible Identity (Layer 2 in Figure 2) framework, it cryptographically fuses workloads/user identities using silicon-level attestation with application-level governance while preserving privacy to create a single, cohesive identity architecture.

Layer 1: Infrastructure Security (The Confidentiality Upgrade Path)

• Confidential Computing (CC) & Trusted Execution Environments (TEEs): Integrates with multi-vendor hardware (e.g., Intel TDX, AMD SEV, and NVIDIA H100 TEEs) to ensure model weights and context remain encrypted in-use, shielding them from privileged admins.
• Integrity for Legacy/Edge: On commodity hardware, AegisSovereignAI uses Keylime and Trusted Platform Module (TPM 2.0) to verify the software stack's Integrity (via Integrity Measurement Architecture (IMA) and Extended Verification Module (EVM)).
• Hardware-Type Binding for Computational Determinism: Restricts model execution to verified silicon types (e.g., NVIDIA H100 vs. A100). This ensures that regulated AI workloads produce consistent, non-drifted outputs by mathematically guaranteeing the specific execution hardware, addressing the "Silicon Lottery" risk. This hardware-rooted provenance supports the reproducibility and documentation principles of MRM and SR 11-7 compliance.

Layer 2: Unified and Extensible Identity (The Provable Bridge)

• Hardware-rooted geo-fenced workload Identity (SPIFFE/SPIRE, Keylime): Binds SPIFFE/SPIRE workload identities to hardware credentials (TPM). An agent cannot execute unless it is on a verified, authorized machine in an authorized geolocation boundary. Privacy-preserving techniques (e.g., Zero-Knowledge Proofs / ZKPs) are used to prove location compliance with regulations without the Enterprise ever having to ingest or store sensitive precise location data.
• Safe Harbor for Bring Your Own Device (BYOD): Securely extend Agentic workflows to unmanaged customer devices by verifying Silicon Integrity on the fly instead of Enterprise Device Ownership. This creates a regulatory Safe Harbor for the Enterprise, proving that data only touched verified hardware without the liability of managing the device itself.
• Combined Human User, Workload, and Device Identity: Combine human user sessions with BYOD/Enterprise device workload (e.g., mobile regulated application) identities to ensure accountability in multi-agent graphs.
• Autonomous Revocation: If a node's hardware state drifts (detected by Keylime), its SPIFFE/SPIRE identity is revoked in real-time, isolating the agent before lateral movement.

Layer 3: AI Governance (Verifiable Logic & Privacy)

• Audit without Disclosure: Privacy-Preserving Compliance. AegisSovereignAI closes the Sovereign Trust Loop by solving the "Black Box" audit problem using privacy-preserving techniques across the complete AI lifecycle: Training Data Ingestion → Inference Data Input → Model Inference System Prompt → Inference Data Output. The Enterprise can provide cryptographically verifiable proof of compliance to regulators (e.g., OCC or ECB) without revealing proprietary prompt logic or sensitive PII.
• Privacy-Preserving Data Ingestion:
• Confidential Transformation: Raw data is ingested into Confidential Enclaves (Intel TDX/AMD SEV). Data is masked and redacted while encrypted in-use, shielded from infrastructure admins.
• Hardware-Rooted Provenance: Uses Fast Identity Online (FIDO)-based attestation to bind data to verified silicon, proving origin is genuine without ingesting persistent hardware identifiers.
• Compliant-Boundary Verification: Privacy-preserving proofs validate that ingested data was collected within a compliant geographic boundary (e.g., Reg-K) without the Enterprise ever touching raw GPS coordinates — transforming a privacy liability into a cryptographic compliance asset.
• Data Minimization Proof: Generates proofs validating that the dataset satisfies GDPR Data Minimization before entering the vector store.
• System Prompt Integrity (Pre-Computed): At deployment, we generate a permanent cryptographic proof that the AI System Prompt includes mandatory safety guardrails (e.g., SSN redaction) and excludes unauthorized directives. This ensures "Compliance-by-Design" without exposing the proprietary prompt text.
• User Prompt Compliance (Batch & Purge): User interactions are processed in real-time while a background process generates aggregated batch proofs. These proofs verify that no user prompts in a given window contained "jailbreak" commands or PII. Once the batch proof is successfully anchored to the enterprise audit log, the raw, high-liability prompts are purged from the system, permanently eliminating PII storage risk.
• AI Output Filtering (Batch & Purge): AI model outputs are verified through real-time Data Loss Prevention (DLP) scanning and content safety checks before delivery. Batch proofs demonstrate that all outputs were properly filtered to prevent hallucinated PII leakage (e.g., fabricated SSNs). Raw outputs are purged after proof generation, ensuring zero retention of AI-generated sensitive data. See the [Privacy-Preserving AI Governance](./docs/auditor-privacy-preserving-ai-governance.md) for the complete three-track verification model.
• Effective Challenge Enablement (SR 11-7): Provides cryptographically-verifiable independent evidence for model validators: proof of execution hardware, data provenance, and governance polic

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: lfedgeai
• Source: lfedgeai/AegisSovereignAI
• License: Apache-2.0
• Homepage: https://github.com/lfedgeai/AegisSovereignAI

Install and usage instructions live in the source repository linked above.