AI guardrails
Prompt injection defences and output controls
What it is
Guardrails are the checks around an AI feature that stop it being misused or saying something it shouldn't. They screen inputs for prompt injection, where hidden instructions in a document or message try to take over the model, redact sensitive data, and check outputs before they reach a user or trigger an action. The OWASP Top 10 for LLM applications is where most security teams start.
Security reviewers now ask specifically about prompt injection and data leaking through AI features. Without a good answer, the AI feature can block the whole deal.
What it takes
About 4–8 engineer-weeks to build in-house, or 2–4 using Lakera, Guardrails AI, NVIDIA NeMo Guardrails.
Answer these first
- Security: What is the worst thing the AI could be tricked into doing?
- Security: Has anyone tried to break it on purpose?
- Engineering: Which content reaching the model could contain instructions written by outsiders?
- Engineering: What can the model access or change, and is that the minimum it needs?
- Legal: Who is accountable when a filtered answer is still wrong or harmful?
- Product: Do users see why something was blocked?
This page works best with JavaScript on. Every answer also has its own address, like /what/scim.