Guardrails are the set of controls placed around an AI model in production to govern what it receives, what it produces and what it is allowed to do.
They work at several levels. On input: detecting prompt injection attempts, filtering prohibited content or personal data. On output: checking the format, blocking toxic content or leaks of sensitive information, making sure the answer stays within the intended scope. For an agent: limiting the tools it can access, requiring human approval before an irreversible action.
They are necessary because an LLM is not deterministic: the same request can produce different answers, and a malicious user may try to hijack it.
They are implemented with coded rules, dedicated classification models (such as Llama Guard), libraries (Guardrails AI, NeMo Guardrails) or the moderation services of model providers and clouds.
Why it matters when hiring
It is a topic for a senior AI Engineer, on the border with application security. In an interview, ask how the candidate tested their guardrails (attack attempts, adversarial test sets) and what they cost in latency. The trap: believing that an instruction in the system prompt is enough to protect an application.