A policy guardrail is a rule that limits, redirects, or blocks risky behavior in an automated system. In cybersecurity, guardrails are used to keep software, AI agents, scripts, and workflows inside approved boundaries. They can enforce least privilege, deny access to sensitive data, require human approval for certain actions, or stop commands that look unsafe.
Guardrails matter because automated systems can move faster than people can react. If an attacker manipulates an AI agent, a guardrail may prevent data exfiltration, unauthorized tool use, or harmful changes to infrastructure. Defenders use them as runtime controls: for example, blocking a suspicious prompt from reaching a model, preventing a bot from running an unapproved action, or limiting what an agent can do even if it is compromised. Good guardrails are precise enough to stop abuse without breaking normal work.