I Built a Guardrails Wrapper to Stop My AI Chatbot Leaking Secrets

A prompt injection detector only checks what goes IN. But what if the input looks completely harmless and the chatbot's own reply leaks the secret anyway? That happened in my red-teaming log, and this project closes that gap. In this video I wrap ZenAssist, my fintech support chatbot from Project 1, in a six-step defence pipeline built on top of Injectguard (Project 2): 1. Rate limit check: lock out sessions with repeated violations 2. Input scoring with Injectguard 3. Enforcement: sensitive topics (internal codes, other customers' data) get a fixed safe reply, and the model is never called 4. The LLM call, only if everything above clears 5. Output filter: catches leaked secrets in the model's own reply, including spaced-out or obfuscated versions 6. Post-enforcement: a backstop if the reply drifts into a sensitive topic Every message produces a full reasoning trail, so nothing is a black box. THE RESULT I replayed the 16 unique attack prompts from my real Project 1 red-teaming log through the pipeline, using the bot's actual historical responses. 5 of them had originally leaked the secret. The wrapper prevented all 5. HONEST LIMITATIONS This is a small, attack-only test set. The enforcement layer is keyword-based, so it won't catch topics it hasn't been told about. And the replay used recorded responses, not fresh live model calls. The full write-up is in the repo. GitHub: https://github.com/sethumak-ai/project3-gu... #AISecurity #LLMSecurity #PromptInjection #Guardrails