Runtime guardrails are technical checks enforced while an AI system is running. They inspect prompts, retrieved context, outputs, and tool calls, then block, rewrite, or limit actions that look unsafe or unauthorized. Unlike policy documents or governance reviews, these controls act in the moment, where an attacker can try to steer a model with malicious instructions or trigger a harmful API action.
They matter because modern AI systems often connect to files, databases, SaaS apps, and agent tools. If prompt injection or data poisoning reaches the execution path, guardrails can reduce leakage, stop disallowed content, restrict tool use, and require human approval for sensitive steps. Common examples include prompt filtering, output moderation, allowlists for tools and domains, rate limits, and rules that detect attempts to exfiltrate secrets or override system instructions. In practice, runtime guardrails provide the enforcement layer that turns AI security policy into real protection.