This research introduces a Root Defence Strategy that intercepts harmful LLM outputs during the generation process itself, rather than after completion.

• Addresses limitations of existing safety methods by operating at the decoding level rather than the prefill level
• Implements adaptive monitoring during token generation to catch harmful content in real-time
• Demonstrates higher effectiveness and robustness against jailbreak attempts and malicious prompts
• Provides a more seamless user experience by avoiding complete response rejections

This approach significantly enhances LLM security by addressing vulnerabilities at their source, offering developers a more reliable method to deploy safe AI systems in production environments.

Root Defence Strategies: Ensuring Safety of LLM at the Decoding Level