LG AI Research: The Concept and Strategy of AI Red Team for Generative AI

As generative AI technology advances, the importance of Responsible AI (RAI) is growing. Accordingly, the AI Red Team, which analyzes problems from an adversary's perspective and prepares countermeasures, plays a key role.

The AI Red Team goes through the following 4-step process to find model vulnerabilities and enhance safety.

• Risk Classification and Identification: Systematically organizes risk factors based on policy, harm type, target, domain, and scenario.
• Attack Strategy Development: Designs and automates attack scenarios that exploit model characteristics based on identified risks.
• Defense Establishment and Validation: Establishes and validates defense methods that can respond to attacks.
• System Operation Improvement: Improves overall AI system operations by reflecting the results of the Red Team.

In particular, the main attack strategies that exploit the characteristics of language models include the following methods.

• Completion Compliance: A method that attaches a positive suffix to induce the model to comply with malicious prompts, leading to harmful responses.
• Instruction Indirection: Exploits the model's ability to follow instructions to bypass safety regulations through specific formats (such as JSON) or complex scenarios.
• Generalization Glide: A method that exploits the model's generalization ability.

LG AI Research is applying this Red Team concept to its enterprise AI agent, ChatEXAONE, to strengthen security requirements and reliability.