Demo library › Walkthroughs

Safe AI Adoption walkthrough

One employee, Alice, and one afternoon. Every AI security control, in the order she runs into them.

~29 minutes · lab or laptop

Before you start

• Setup done, including Authenticate the MCP servers. Connect the agent before you start, or it spends the first agent step asking for a connection.
• Signed in as in the browser and in Goose. Use Goose: it shows the audience why a request was refused.
• Run the whole walkthrough in log mode first, then again in block mode. The second pass is quicker.
• Start each chapter in a new Goose chat.
• If the agent starts looking for another way in after a refusal, stop it: "stop, just tell me what was refused". That is what agents do when a door closes, and worth saying out loud.

Chapters

1. Customer data pasted into a public AI The company's AI decisions enforced - ChatGPT inspected, Gemini isolated, the rest redirected - and a customer list blocked on its way into ChatGPT. · ~6 min · Lab and laptop
2. An AI agent reads what the app hides The calendar app shows Alice nothing confidential, but its MCP tool returns the company's confidential meetings - until Gateway DLP blocks the response. · ~6 min · Lab and laptop
3. Personal data pasted into the company assistant Alice pastes her own HR record into the company assistant. AI Gateway blocks the prompt before it reaches the model. · ~4 min · Lab and laptop
4. Asking the assistant to help commit a crime Alice asks the company assistant to help her take money from the company. AI Gateway guardrails refuse it before any model is called. · ~3 min · Lab and laptop
5. An AI agent asks for data Alice cannot see Ledger's tools are not offered to Alice's agent at all, and the CRM's over-sharing tool is blocked by Gateway DLP. · ~5 min · Lab and laptop
6. Show the auditor what happened Every decision from the other chapters in the logs, the matched data and prompts decrypted with the company's own key, and a user risk score. · ~5 min · Lab and laptop

Afterwards

Open both consoles, filtered to Alice:

• Zero Trust → Insights → Logs → Gateway: the blocked ChatGPT prompt, the claude.ai redirect, and the blocked WorkBox and Pipeline tool responses. There is nothing for Ledger, because its tools were never offered.
• AI Gateway → Logs: every model call, including the blocked prompt and the guardrail refusal.

End on this: none of the applications logged anything, because they all behaved correctly. Every control sat between Alice and something else. The real incidents page has the published cases behind each chapter.