Method for Security Control of Prompts and System Behaviors Executed by AI Gateway

Abstract

The present invention relates to a security control method for prompts and system behaviors executed by an AI gateway in an environment comprising an AI service instance, a generative AI model, and an AI gateway positioned between the AI ​​instance and the generative AI model. A security control method according to the present invention comprises: a first step in which an AI gateway receives a prompt input by a user through an AI service instance; a second step in which an AI gateway analyzes the semantic and structural characteristics of the received prompt; a third step in which an AI gateway analyzes system behavior executed during the prompt processing process; a fourth step in which an AI gateway calculates a risk level based on the analysis results of the second and third steps; a fifth step in which an AI gateway transmits the input prompt to a generative AI model as is, transmits it modified, or blocks it based on the risk level calculated in the fourth step; a sixth step in which an AI gateway analyzes a response output by a generative AI model; a seventh step in which a response risk level is calculated based on the analysis results in the sixth step; and an eighth step in which a security operation is executed based on the risk level calculated in the seventh step, such as transmitting the response to an AI service instance as is, at least masking a part of the response, summarizing the response, or blocking the response.