Why support chatbots must be designed around what they refuse to answer

A developer building a support chatbot for a product landing page found that conversational AI systems are vulnerable to prompts designed to extract system instructions or bypass intended restrictions. Rather than using keyword-based filters in application code, the team embedded explicit refusal policies directly into the LLM's system prompt, allowing the model to judge requests by category rather than by specific banned words. This approach leverages the model's own language understanding to handle rephrased or unanticipated inputs that static filters would likely miss. The design also governs how refusals are worded, favouring brief, uninformative responses to avoid giving users clues about where the boundaries lie. The key takeaway is that a chatbot's reliability depends as much on how it handles prohibited requests as on how well it answers legitimate ones.