DEEPALIGN is a robust defense framework that fine-tunes Large Language Models to progressively detect and detoxify harmful content during the generation process.

• Addresses a critical security gap where traditional safety methods fail by examining content only in initial generation steps
• Introduces progressive answer detoxification that continuously monitors and adjusts output toxicity
• Demonstrates superior performance against multiple jailbreak attack types compared to baseline approaches
• Provides a computationally efficient solution that balances security with performance

This research significantly advances LLM security by targeting the dynamic nature of jailbreak attacks, offering organizations a practical framework to deploy safer AI systems without compromising functionality.

Align in Depth: Defending Jailbreak Attacks via Progressive Answer Detoxification