A gatekeeper model is a lightweight AI model placed in front of a larger model or agent to quickly screen user input. Its job is to decide whether a prompt looks safe enough to pass on for full processing, or whether it should be blocked, flagged, or sent for deeper review. Because it is designed for speed, it usually has less context and weaker reasoning than the downstream system.
Gatekeeper models matter because they are often the first line of defense in AI security. Attackers try to defeat them with prompt injection, obfuscated instructions, or ordinary-looking prose that hides harmful intent. If the gatekeeper is too permissive, malicious content can reach a more capable model with tool access, data access, or execution privileges. Defenders use gatekeepers as one layer in a broader control stack: combine them with instruction/data separation, output monitoring, sandboxing, and least-privilege access so that bypassing the first filter does not automatically lead to compromise.