LLM-Firewall

A security gateway that sits between users and an LLM. Type a prompt to see how each detection layer scores it, what the policy decides, and what the output scanner redacts from the (mock) model's answer.

This public demo runs Layer A (heuristics) and Layer B (TF-IDF + logistic regression) only. Layer C (fine-tuned transformer) is too large for a serverless function and runs only locally / in Docker. The "LLM" here is a built-in mock that deliberately misbehaves (leaks its system prompt, returns fake PII) so the output-side defenses have something to catch. No real LLM or API key is used. Requests are rate-limited.

Prompt