Hardware-enforced private AI infrastructure.

We build private AI systems and AI agents that run on hardware you control, with no path for your data to reach the outside world. They’re designed for organizations that must meet CMMC 2.0, HIPAA, and other strict data-protection rules.

The containment stack as privilege rings Concentric rings from ring 3, apps and agents, through circuit breakers, sanity monitoring, kernel and access control, enclave isolation, and firmware, down to the hardware core where your data lives. R3 apps & agents R2 circuit breakers R1 sanity monitor R0 kernel & access R−1 enclave R−2 firmware R−3 data
The rings show where our safeguards sit. AI apps and agents run on the outside, and each layer beneath them, down to the hardware, adds a control the AI can’t reach or switch off. The sections below follow the same rings inward.

R&D labled by research scientists and engineers

Decadesacross real-time embedded systems, neuromorphic computing, machine learning, systems engineering, and security

0public AI services handling your data

4 frameworksCMMC 2.0, HIPAA, FERPA, ITAR

R3Threat matrix

Your staff are already using AI, and it’s creating risk.

Most of that risk falls into three areas. For each one, we put the safeguard underneath the AI, where it can’t be switched off or talked around.

T1Data leaking into public AI

How it happens

Staff paste contracts, source code, and patient notes into public chatbots. Those services may keep, log, or train on what they receive, under terms no one at your organization reviewed.

What it costs

Proprietary and regulated data leaves your organization, with no record of where it went.

How we contain it

Private AI that runs on hardware you control, with no connection to the outside.

T2AI agents acting on bad output

How it happens

AI agents can now call other systems, update databases, and chain tasks together, all based on model output that is sometimes simply wrong.

What it costs

One made-up value can turn into a deleted table, a bad wire transfer, or an email to the wrong customer.

How we contain it

Circuit breakers that check every action an agent proposes against fixed rules before it runs.

T3Compliance violations

How it happens

Commercial AI services sit outside the systems your auditors approved, and outside the agreements your data requires.

What it costs

Failed CMMC assessments, lost defense contracts, HIPAA penalties, and breach notifications.

How we contain it

Systems designed from day one around NIST SP 800-171, HIPAA, FERPA, and ITAR.

R2Architecture

Every action an AI agent takes is checked before it runs.

Most AI guardrails are written instructions, and AI models don’t always follow instructions. Our safeguards are software that sits between the AI and your systems. If a request fails a check, it doesn’t run.

What happens when an AI agent tries to do something

  1. 1

    Request

    The agent asks to take an action, such as calling a system or changing a record. Nothing has happened yet.

  2. 2

    Format check

    The request must match an exact, expected format. Unknown tools and malformed requests are rejected.

  3. 3

    Limits

    Hard limits you set: which systems, which records, how many at once, and how much it can cost.

  4. 4

    Second opinion

    A separate AI model reviews the request and stops it if the agent has drifted off course.

  5. 5

    Human approval

    High-impact actions wait for a person to approve them with a hardware security key.

  6. 6

    Run and record

    The action runs inside your network, and every step is written to the audit log.

Nothing leaves by default. There’s no outbound connection for a prompt, a document, or usage data to escape through.

Air-gapped private systems

Dedicated hardware in your building, or isolated accounts in AWS GovCloud (US) and Azure Government. No connection to public AI services, and no usage data sent back to any vendor.

Circuit breakers for AI agents

Every request an agent makes is checked against fixed rules before it runs. However confident the AI sounds, it can’t argue its way past them.

Behavior monitoring

A second layer compares results with expected, known-good behavior. If an agent starts to drift, it drops to read-only until someone reviews it.

Private document search

Your documents are indexed and searched on your own hardware. Each person’s access rights carry through, so the AI only sees what that person is allowed to see.

Signed approvals

Sensitive actions need a person’s sign-off with a hardware key, such as your existing PIV or CAC cards. Each approval covers one exact action, can’t be reused, and records who approved what, and when.

Records your auditors can use

Prompts, searches, blocked actions, and approvals are all logged inside your network, in a format built for audit review.

R1Deployment

In your building, or in a government cloud you own.

The safeguards are the same either way. What changes is where the hardware lives, and whether its separation from the outside world is physical or enforced in software.

On-premises equipment

Where it runs
In your facility, on dedicated GPU servers that we specify, install, and secure.
Isolation
Physical. It can run completely offline, with no network connection out of the building.
AI models
Open-weights models sized to the hardware, from small edge devices to multi-GPU racks.
Best for
Controlled unclassified information (CUI), ITAR technical data, patient records, and anything that must never leave the building.

Private government cloud

Where it runs
In your own accounts in AWS GovCloud (US) or Azure Government (GCC High).
Isolation
Enforced in software. Outbound traffic is blocked, and confidential-computing hardware (AMD SEV-SNP, Intel TDX and SGX) keeps data encrypted even while it’s being processed.
AI models
The same open-weights models, running in your account and shared with no one.
Best for
Teams spread across locations, or workloads that grow and shrink, that still need a FedRAMP High authorized cloud.

Keeping an offline system current

  1. 1

    Download outside

    Updates and new model versions are downloaded on a separate, internet-connected system.

  2. 2

    Verify

    Each file is checked against the publisher’s cryptographic hash and a signed list of approved files, then recorded with its source and version.

  3. 3

    Transfer in

    Files cross into the enclave on reviewed media or through a one-way transfer device.

  4. 4

    Verify again

    Inside the enclave, the hashes are checked again before anything is installed.

Each model stays locked to a tested version. A new version goes through change control with its own testing, so the AI’s behavior never changes without review.

R0Compliance

Built for the rules you have to follow.

Public AI services break these rules for the same reason: your data leaves systems you’re required to control. A private system keeps it inside, and we document which requirements each part of it supports.

How a private AI system supports CMMC 2.0, HIPAA, FERPA, and ITAR
Framework Who it applies to Where public AI falls short How a private system helps Controls covered
CMMC 2.0 Level 2NIST SP 800-171 Defense contractors and subcontractors that handle controlled unclassified information (CUI). Sending CUI to a commercial AI service takes it outside the systems your assessor approved. The AI, its search indexes, and its logs stay inside those systems, with built-in records of who accessed what, what happened, and who signed in. ACAUIASC
HIPAA Security Rule45 CFR 164, Subpart C Healthcare providers, health plans, and their business associates that handle electronic patient records. Patient data ends up with an outside AI company, often without a business associate agreement or an audit trail. The AI runs on your hardware, with the access controls, audit logs, integrity checks, and encryption in transit that the rule requires. §164.312
FERPA34 CFR Part 99 Schools and universities that receive U.S. Department of Education funding. Student records are shared with an AI vendor the school doesn’t directly control. Student records, research, and the search indexes built from them stay on systems the school controls. AccessDisclosure
ITAR22 CFR 120–130 Manufacturers and integrators that handle defense articles and technical data. Processing technical data on overseas servers, or where foreign persons can reach it, can count as an export. Hardware located in the U.S., access limited to authorized U.S. persons, and no data path out of the country. LocationAccess

Where the CMMC Level 2 requirements land

Parts of a private AI system mapped to NIST SP 800-171 requirements
Part of the system What it does NIST SP 800-171 requirements
Outbound-blocking network boundary Nothing leaves except a few specifically approved connections, and the AI has no route to the internet. 3.13.13.13.63.1.33.1.20
Directory sign-in with MFA Every person and service account signs in through your directory with multifactor authentication. 3.5.13.5.23.5.33.1.1
Permission-aware document search Each search is limited to what the person asking is already allowed to see. 3.1.13.1.23.1.5
Circuit breaker Checks format, permission, and limits before any action runs. The agent holds no credentials. 3.1.23.1.33.1.53.1.7
Signed approvals High-impact actions are signed by a named person with a hardware key. 3.3.23.1.73.5.2
Audit log Prompts, searches, decisions, approvals, and results, recorded in a log that can only be added to. 3.3.13.3.23.3.83.3.9
Encryption and key management FIPS-validated encryption in transit and at rest, with keys held by you. 3.13.83.13.103.13.113.13.16
Confidential computing (cloud option) Memory stays encrypted while data is being processed. 3.13.113.13.16

A CMMC Level 2 assessment covers all 14 requirement families across your organization, including configuration management, media protection, personnel security, physical protection, and incident response. The AI system is built to fit inside that program. Its design documents, boundary diagram, and control mapping go straight into your System Security Plan, and its audit log lets an assessor trace any AI-assisted action to the person who requested it, the rules that allowed it, and the person who approved it.

AC, AU, IA, and SC are the NIST SP 800-171 control families for access control, audit and accountability, identification and authentication, and system and communications protection. This mapping shows how our systems support technical requirements; certification still comes from your organization and its assessors. Sovereign AI Forge is not a CMMC Third-Party Assessment Organization (C3PAO) and doesn’t provide legal advice.

R−1The 30-day audit

Start with a 30-day audit.

The Sovereign AI Infrastructure Audit is a fixed 30-day engagement. We find out how AI is actually being used in your organization, where data is getting out, and what a private replacement should look like.

  1. Week 1

    Inventory

    We find every AI tool in use, approved or not, including chat tools, browser extensions, coding assistants, and connected agents, along with the data each one touches.

  2. Week 2

    Assess

    We trace how data is getting out, whether through prompts and uploads, usage data sent back to vendors, vendor retention and training terms, or agents acting without review.

  3. Week 3

    Design

    We design the private replacement, including where it runs, which AI models it uses, which safeguards it needs, and how activity gets recorded.

  4. Week 4

    Blueprint

    We walk your leadership and security team through what we found, the proposed design, and a phased plan they can budget for.

  5. After

    Build and support

    If you’d like, we build the system in gated phases. A read-only pilot comes first, and agents get access to real systems only after attack testing shows the safeguards hold. We stay on as the engineers who support it.

R−2Whitepaper

Read the technical details.

Technical whitepaper

Hardware-Enforced Circuit Breakers & Air-Gapped Multi-Agent Containment for CMMC Level 2+ Environments

AI agents are arriving in the defense supply chain faster than the controls meant to contain them. CMMC 2.0 and NIST SP 800-171 require controlled unclassified information to stay inside an assessed boundary, yet most agent frameworks assume a public AI service and unlimited access to tools.

This paper describes an architecture that resolves the conflict. It combines open-weights models on hardware the contractor controls, circuit breakers that check every agent action before it runs, and cryptographic human approval for the actions that matter. Each part is mapped to the NIST control families for access control, audit, identification and authentication, and system and communications protection.

Contents

  1. How public AI services leak data
  2. Inside a private AI enclave
  3. Circuit breakers and sanity checks for agents
  4. NIST SP 800-171 and CMMC control mapping
  5. Implementation roadmap
Request the whitepaper
Sovereign AI Forge logo: a shield split into two bracket-shaped halves, holding a flame with the four-point AI sparkle at its heart

R−3About us

Built by the people you’ll talk to.

Research scientists & systems engineers

Sovereign AI Forge is an engineering firm led by research scientists and engineers. Together they bring decades of experience in real-time embedded systems, neuromorphic computing, machine learning, systems engineering, and security, along with doctoral-level research in computer science. Designing for redundancy and planning for failure come with that experience.

We build every system ourselves, from the hardware and the private environment to the open-weights AI models that run inside it and the safeguards that control what those models can do. We don’t resell someone else’s AI service or put our name on a public one.

When your security team has questions about encryption keys, network diagrams, or audit logs, they get answers from the engineers who designed the system, from the first briefing through the years it runs.

FAQCommon questions

What security teams ask us first.

Do you resell OpenAI, Anthropic, or other cloud AI services?

No. We don’t resell, repackage, or connect you to public AI services. We build private systems that run open-weights AI models on hardware you control, either on-premises or in a government cloud account you own. Your prompts, documents, and results never pass through an outside AI company.

Can we use ChatGPT, Copilot, or other public AI tools with CUI?

Generally not, unless the specific service sits inside your assessed boundary. DFARS 252.204-7012 requires any cloud service that stores or processes covered defense information to meet security requirements equivalent to the FedRAMP Moderate baseline, and your System Security Plan has to account for it. A general-purpose public AI service used outside that boundary doesn’t meet the bar, whatever its privacy settings. Some providers offer government-cloud versions that may qualify, and your assessor should confirm any specific service. A private system sidesteps the question by keeping the AI inside the boundary you already control.

What does “air-gapped” mean in your deployments?

On-premises, it means the AI system has no network connection to the internet at all. The models, search indexes, and logs live on hardware inside your building, and updates arrive through a controlled, reviewed transfer. In AWS GovCloud or Azure Government, the separation is enforced in software instead: the system runs in your own account, outbound traffic is blocked by default, and confidential-computing hardware keeps data encrypted even while it’s being processed.

How do you keep an air-gapped AI system up to date?

Through a controlled transfer. Updates and new model versions are downloaded on a separate internet-connected system, checked against the publisher’s cryptographic hash and a signed list of approved files, and recorded with their source and version. They then cross into the enclave on reviewed media or through a one-way transfer device, and the hashes are checked again before anything is installed. Each model stays locked to a tested version, so a new one goes through change control with its own testing.

Can private AI help us meet CMMC Level 2?

Yes, if it’s designed for it. A private system keeps CUI inside your assessed environment and produces the access, activity, and sign-in records an assessor looks for. We map each part of the system to the NIST SP 800-171 controls it supports. Certification itself comes from your assessment by a C3PAO, the independent assessors authorized to certify CMMC; our job is to make that assessment straightforward.

Which NIST SP 800-171 requirements does a private AI system help satisfy?

Mainly requirements in four families. In access control: 3.1.1, 3.1.2, 3.1.3, 3.1.5, 3.1.7, and 3.1.20. In audit and accountability: 3.3.1, 3.3.2, 3.3.8, and 3.3.9. In identification and authentication: 3.5.1 through 3.5.3. In system and communications protection: 3.13.1, 3.13.6, 3.13.8, 3.13.10, 3.13.11, and 3.13.16. The compliance section shows which part of the system supports each one. A CMMC Level 2 assessment covers all 14 families across your whole organization, so the AI system is built to fit inside your existing program and supply evidence for it.

Which models do you run?

Open-weights AI models, chosen for the task and sized to the hardware, from compact models on small edge devices to large models on multi-GPU servers. The models are stored and run inside your environment, and each one is locked to a tested version so its behavior doesn’t change without warning.

What is a deterministic circuit breaker?

It’s a checkpoint between an AI agent and the systems it can reach. When an agent wants to do something, such as call another system, change a database, or send a message, the circuit breaker first checks the request against fixed rules: which tools are allowed, which records, how many, and at what cost. Anything outside those limits is blocked, and high-impact actions wait for a person to approve them with a hardware security key. Because the rules are written in software, the AI can’t talk its way past them.

What is indirect prompt injection, and why don’t prompt guardrails stop it?

AI agents read content from outside your organization, such as documents, web pages, and email. An attacker can hide instructions in that content, and the model may follow them, for example by sending data somewhere using tools it was legitimately given. Telling the model to ignore such instructions helps, but it can’t be relied on, because the model being manipulated is the same model being asked to refuse. That’s why our safeguards sit outside the model. An injected instruction can change what the agent asks for, but it can’t change the circuit breaker’s rules, the credentials it holds, or the network paths that exist.

How are high-impact AI actions approved?

A named person signs each one with a hardware key, such as your existing PIV or CAC cards or a FIPS 140-validated security key. The approver sees a plain-language summary and the exact action that will run, and the signature covers that exact action, a one-time value, and an expiry time. If anything about the action changes after signing, or the approval is reused or expired, the circuit breaker refuses it. The audit log then shows who approved each action and proves the action that ran is the one they approved.

What happens if an AI agent starts behaving differently?

Behavior monitoring watches each agent over time: how often its requests are refused, which tools it uses, how its results compare with a reviewed reference set, and how often the second-opinion check raises a flag. If a measure moves outside its approved range, the agent drops to a restricted mode, such as read-only, until someone reviews it. A rising number of refused requests is often the first sign that an agent is being manipulated or that a workflow needs redesign.

Do we have to buy hardware?

Not necessarily. On-premises equipment gives the strongest isolation and suits CUI, ITAR technical data, and patient records that must stay in the building. If you’d rather not run hardware, the same safeguards can run in AWS GovCloud (US) or Azure Government accounts you own. The 30-day audit recommends the option that fits your data and your risk.

What does the 30-day audit deliver?

You get a list of the AI tools already in use across your organization, including ones no one approved; a ranked assessment of where data is leaking or could leak; a design for a private replacement; a map of that design to the rules you follow; and a phased roadmap with cost estimates. Many clients have us build the system afterward, but the blueprint is yours either way.

Where are you located, and who do you work with?

Sovereign AI Forge, LLC is based in Port Charlotte, Florida. We work with defense prime contractors and subcontractors, federal systems integrators, healthcare networks, universities, and businesses across the United States. In Southwest Florida, from Sarasota to Naples, we work with businesses on-site as Suncoast Sovereign AI, our regional team at suncoastsovereignai.com. Hardware installations and executive briefings are done in person.

CoreStart here

Request a technical audit.

Tell us what you need to protect and which rules you have to follow. We’ll get back to you to set up a confidential technical briefing with one of our engineers.

Please don’t include CUI, PHI, or other controlled information in this form.

Prefer to talk it through? Call our engineers at (941) 436-0420.