Demo library › Chapter 8
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.
The problem
Alice needs a messy customer list tidied, and ChatGPT is right there. The company has approved one assistant, has another under review, and has not approved the rest.
Lands with: CISOs and anyone writing an AI policy
Before you start
- The client signed in as
. - Run once in log mode, then again in block mode.
Run it
What the company has decided
| Do | Say |
|---|---|
| Zero Trust → Team & Resources → Application Library, filtered to Artificial Intelligence. | Each AI service has a status: ChatGPT approved, Gemini in review, the rest unapproved. |
| Traffic policies → Firewall policies → HTTP: Prevent sensitive data reaching public AI, Isolate applications in review, Redirect non approved AI. | Approved services are inspected, services in review are isolated, and unapproved ones are redirected to the approved options. |
Whybackground, if you are asked
The rules match on the review status, not on a list of hostnames. An AI service nobody has evaluated yet is handled the same way without anyone updating a list.
In the browser
| Do | Say |
|---|---|
| Open Nexus → Marketing → Summit invite list (pasted from CRM - needs tidying) and copy the list. | Ten customers with names, titles and mobile numbers, and notes about who is a churn risk. A colleague pasted it into the wiki and never cleaned it up. |
Open claude.ai. | Redirected to a page offering the two approved options: ChatGPT for non-sensitive data, and the company's own AI for sensitive data. |
Open gemini.google.com, ask anything, then try to copy the answer. | It works in a remote browser, but nothing can be copied or pasted. |
Open chatgpt.com and paste the list with: Format this into a table with columns for name, title, company, email and phone: | In log mode, a neat table comes back - and ten customers' mobile numbers have left the company with no record. In block mode, the prompt is blocked and the client shows a notification. The site itself still opens. |
| Paste the same request into the company's own assistant in Goose. | It works. The company's assistant is allowed customer data, because working with customers is a marketer's job. |
Whybackground, if you are asked
Alice has not done anything she would think twice about. She used the tool in front of her for what it is for.
Blocking the site is what most organisations try first, and it fails: people use their phones. Blocking the data while leaving an approved path open is what changes behaviour. The work moves somewhere the company can see, log and attribute.
What the audience should see
- claude.ai redirected, Gemini with copy blocked, and the ChatGPT prompt blocked.
- The company assistant answering the same request.
Where the evidence is
Gateway HTTP logs: the blocked ChatGPT request, and a separate redirect entry for claude.ai. Decrypt payload log shows the prompt.
Known limitations
- Prompt inspection covers ChatGPT by name. Another approved AI service would need its own rule.
Part of: The whole of Cloudflare One · Safe AI Adoption walkthrough · The full tour