All demo scripts

Pasting a colleague's details into a chat

The same DLP profiles that guard MCP traffic also inspect the prompt itself. This is the everyday case: nobody is attacking anything, someone is just trying to get their work done faster.

Set the scene

Delta has a spreadsheet open — an export someone sent her for a mail merge. She wants the assistant to tidy it up, so she pastes a few rows into the chat. She is not exfiltrating anything; she is doing what the tool is for.

This is the shape of most real incidents, and it is the one an agent-focused demo usually misses: the data never came from a tool, so nothing at the MCP layer is involved.

Show the web UI first

Show that the data is perfectly ordinary in its own context: open and look at Delta's own record, which she is entitled to see. The problem is not where the data lives, it is where it is about to go.

Now ask the agent

Expected tool calls: 0 - nothing is called, the prompt itself is the problem. Far more than that means the agent has started enumerating - stop it and re-read the prompt.

Variations that work the same way:

• Draft a friendly email to this person confirming their details: [paste a row including an address and a national identifier]
• Summarise this HR note for me: [paste a case note mentioning a performance improvement plan]

What happens with no protection

The prompt goes to the model, and with it a home address, a date of birth and a national identifier. Depending on the provider it may be retained, logged, or used for training. The employee has no idea any of that happened, and neither do you — there is no record that this prompt was different from any other.

What happens with protection deployed

AI Gateway applies the same four DLP profiles to the request body. Employee PII matches the national identifier and the mailing address — both via Cloudflare's own predefined detection entries — and the request is blocked before it reaches the model:

{
  "error": {
    "code": 2011,
    "message": "Request blocked due to data loss prevention configurations"
  }
}

The gateway checks REQUEST and RESPONSE, so a model repeating an identifier back — from an earlier turn, or from its own invention — is caught on the way out too.

Watch the tool stream while it runs

Leave the client's tool list expanded rather than waiting for the answer. Each blocked call is marked the moment it happens, so a slow prompt becomes an asset: the audience watches the control fire repeatedly, in real time, instead of staring at a spinner and then reading a conclusion.

Execute  const employees = await tools["ai-demo"].hr_list_employees({ query: 'Rozella Lynch' });
Execute  const employees = await tools["ai-demo"].hr_list_employees({ query: 'Rozella' });
         ! Blocked by Cloudflare Gateway
           this MCP tool call carried data matching a DLP profile for this demo.
           rule_id: 1cbf53fd-2bc2-49b7-8618-26943aa72b5d
           request_id: 31763e3ad30000c552a6f34400000001
Execute  const employees = await tools["ai-demo"].hr_list_employees({ query: 'Lynch' });
         ! Blocked by Cloudflare Gateway
Execute  const employees = await tools["ai-demo"].hr_list_employees({});
         ! Blocked by Cloudflare Gateway

Point at three things in that stream:

• The agent retries. Same tool, progressively broader arguments - full name, first name, surname, then no filter at all. Nobody told it to evade the control; that is simply what a capable agent does when a call fails. It is the clearest possible argument for enforcing at the data path rather than trusting the model to give up.
• Every attempt is blocked, not just the first. The control does not tire and does not care how the question is phrased.
• The request_id is a bridge. Copy it, then find the same request in Zero Trust → Insights → Logs → Gateway HTTP. Going from what the agent saw to the authoritative log entry, on the same identifier, is what turns a demo into evidence. The rule_id names which of the four policies fired.

Then expand the agent's own reasoning

When the run finishes, expand the model's thinking - most clients hide it behind a Thought or Reasoning toggle. The model narrates the control working, in its own words. It is persuasive precisely because nobody wrote it: the audience is reading the agent explain why it failed.

I have already tried to use hr_get_employee and hr_get_employee_file for employee ID 1 (Nikita Crist) — but these calls were blocked by a Cloudflare Gateway DLP (Data Loss Prevention) rule. […] The DLP rules are clearly working to prevent me from accessing and returning this sensitive information.

Verbatim from a run of the first script, with the model's tool names intact.

Three things to draw out of whatever your run produces:

• The tool names. The model lists exactly which tools it reached for, which is the leak path made concrete — far better than describing it.
• Who stopped it. It names Cloudflare Gateway and DLP. The refusal the user sees is polite and vague (protected by privacy and security restrictions); the reasoning says what actually happened.
• What it tried next. A blocked agent does not stop, it re-plans. Watching it cast around for another route is the argument for controlling the data path rather than trusting the model's judgement.

Where to show the evidence

• AI Gateway → Logs: the blocked request, with the DLP policy that matched. Contrast it with the Gateway HTTP logs used by the other demos - different product, different log, same profiles.
• Zero Trust → DLP profiles: show that Employee PII is one profile used in two places, so a detection added once applies to both paths.

Choosing the paste

Use data from the seeded directory rather than inventing some, so a sceptic can verify the record exists in afterwards. Any row with an address and a national identifier works; the profile matches the identifier formats, not the names.