LG AI Research: The Concept and Role of AI Red Team

As generative AI technology advances rapidly, the importance of Responsible AI(RAI) for safely building and deploying systems is growing. Accordingly, AI Red Team, which originates from military strategy that simulates problems from an adversary's perspective to prepare countermeasures, is drawing attention.

AI Red Team detects new security threats such as Prompt Injection and Data Poisoning, and prevents bias and the occurrence of harmful content. In particular, considering the characteristics of generative AI, which uses probabilistic generation methods, it builds multi-layered defense strategies through various tests and analyses, enhancing the system's consistency and reliability.

The AI Red Team process largely consists of four stages:

• Risk Identification and Classification: Defining and classifying risks the model may have
• Attack Strategy Development and Automation: Building attack scenarios for identified risks
• Defense System Construction and Verification: Designing countermeasures to respond to attacks
• System Operation Improvement: Reflecting the Red Team's results in actual system operations

LG AI Research applied these AI Red Team techniques to ChatEXAONE, an enterprise AI agent released last August. Through this, it is strengthening security requirements and reliability, providing users with more responsible AI services.