Defense AI: How It Works (Architecture and Request Flow)

Last updated: November 11, 2026

Defense AI gives your team Claude Desktop, Claude Code, Codex and OpenCode, connected to AI models that run in your company's own cloud. Prompts and responses travel from each person's computer to your company's own gateway in your Azure subscription, then to your models. They don't pass through Atomus.

The three parts

  • The gateway. Each company gets its own gateway, deployed in its own Azure subscription. It checks each person's key, applies your budgets and rate limits, and sends the request to your models.

  • The models. OpenAI models run on Microsoft's Azure OpenAI service in your Azure subscription. If your company uses the optional Claude add-on, Claude models run on Google Cloud Vertex AI in a Google Cloud project your company owns. Claude requests leave your Azure environment: the gateway sends them over an encrypted connection to Google's US Vertex AI endpoint.

  • The admin portal. Atomus hosts the Defense AI admin portal (ai.atomuscyber.us). Your admins use it to connect your cloud account, choose models, set budgets and rate limits, and manage connectors and users. The portal stores settings and daily usage totals. It is not in the path of prompts or responses.

Your company pays its cloud provider directly for model usage. No AI processing runs in an Atomus cloud account.

How a request flows

  1. A person sends a prompt from Claude Desktop, Claude Code, Codex or OpenCode. The tool is set up with their personal Defense AI key.

  2. The request goes over HTTPS to your company's gateway in your Azure subscription.

  3. The gateway checks the key, applies your limits and routing rules, and sends the request to the model in your Azure subscription or your Google Cloud project.

  4. The response comes back the same way.

The admin portal only manages settings and reads daily usage totals from the gateway.

What the gateway keeps

  • Prompts and responses pass through the gateway's memory while a request is being handled. The gateway doesn't write them to disk or to its database.

  • Usage totals are stored: request counts, token counts and estimated cost.

  • Response caching is off by default. If an admin turns it on, cached responses live only in the gateway's memory for a short time and are never written to disk.

This covers Defense AI and your gateway. Microsoft and Google run their own abuse monitoring on the models, which can keep flagged content until your company is approved for an exemption. See Defense AI: Data Retention, Who Sees Our Data, Is it Used for Training.

Who does what

Role

What they do

Company admin

Runs cloud setup, chooses models, sets budgets, manages connectors and users

Employee

Signs in on your company's Connect your tools page with their Microsoft work account, gets a personal key, and connects their tools

Atomus support staff

Provide technical support, maintain the model catalog and pricing, and roll out gateway updates

Atomus read-only staff

View setup status for support, with no ability to change your configuration

Why does Claude Desktop show a GPT model?

Claude Desktop and Claude Code can use any model your gateway offers. If your company's models are OpenAI models on Azure, Claude Desktop shows those model names. You keep the Claude Desktop and Claude Code apps; the model behind them is whichever one your admin deployed.

Can we keep using commercial AI for non-CUI work?

Defense AI doesn't require you to stop using other tools. Many companies keep a commercial AI plan for work their policy allows there and use Defense AI for CUI. Your company's policy decides which work goes where. In Claude Desktop, check that the bottom-left corner says Defense AI before you work with CUI.

Related articles: Defense AI: Where Your Data Is Processed (US Data Centers and FedRAMP) · Defense AI: Can We Use It for CUI or ITAR Data? · Defense AI: Data Retention, Who Sees Our Data, Is it Used for Training · Defense AI: Security Controls