LLM Firewall: Rule-Based Jailbreak Detection and Limits

Evaluate prompt-injection controls at an observable input/output boundary using hostile instructions and legitimate task examples.

LLM Firewall applies local rules and structural checks to supplied prompt, output, retrieval, and tool events. Its core checks do not require a remote LLM classification request.

A rule-based detector does not guarantee zero-day jailbreak detection. Run regression cases and a representative benign corpus; enforce tool permissions independently of model text.

Check collector wiring, the current event schema, and the enforcing component before relying on a response. Preserve the original event and verify the result on your actual deployment. An emitted alert does not prove that the initiating operation was stopped.