4.11 — Security Quiz 📊

Match each problem to its countermeasure.

Problems

• A. The AI starts replying to an aggressive customer… with insults.
• B. The AI is about to offer a €500 refund to a customer.
• C. You want to know how much this AI cost you this week.
• D. A PDF sent by a “customer” contains a hidden instruction ordering the agent to exfiltrate your customer database.
• E. The agent retries a failing connection, over and over, all night.

Then:

F. In a Make/n8n scenario that sends customer emails, where do you place the HITL point?

G. Your agent only needs to read an Airtable table. Which principle do you apply to its permissions?

H. True or false: with a no-code agent, you have no trace of what it did.

👉 Show answers

• A → Guardrail : filters the tone / blocks toxic outputs before sending.
• B → Human-in-the-Loop : the AI freezes the refund and waits for your “Approve” click.
• C → Observability : the traces dashboard to count tokens spent.
• D → Indirect injection : a guardrail on incoming content + least privilege (the agent isn’t allowed to export the database) + break the lethal trifecta.
• E → Stopping rule : iteration cap + no-progress detection.
• F → Right before the sensitive/irreversible action (sending the email), via an approval step that pauses the scenario.
• G → Least privilege : read-only access, without the right to delete or export.
• H → False. Make/n8n keep an execution history and Airtable a revision history — a first native layer of traces.