Guardrail Stack Builder
Spec your app and get a defense-in-depth guardrail stack — six ordered layers from input filter to tool-call allowlist — with a concrete component chosen per layer, its rationale, its known bypass, the latency it adds, and a copy-pasteable starter config. Export the whole thing as a Markdown runbook.
Runs entirely in your browser. Recommendations are heuristic engineering guidance, not a substitute for threat modeling and red-teaming.
Why these layers, in this order?
Guardrails work as defense in depth: no single control stops a determined attacker, so each layer assumes the previous one was bypassed. The order matters — cheap deterministic checks run first, expensive model-based classifiers only on what survives, and the tool-call allowlist sits last because it is the only thing that reliably contains an injection that already reached the agent.
Background reading on LLM guardrails and the offensive Attack Technique Atlas these layers map onto.