A defensive prompt can make an assistant withhold information it already has. In Anthropic'sprompting workshop, a fictional mobile-support bot receives a customer's legacy-plan allowance but redirects the customer elsewhere. Its instructions emphasize avoiding incorrect plan details and sending customers to a website.
The presenter replaces that instruction with a clear rule that the supplied customer record determines the allowance, then reports the relevant tests passing. The example shows why prompt maintenance needs tests for useful answers as well as mistakes: an old protective instruction can become the reason a capable model fails the task.