This research introduces LongSafety, the first comprehensive benchmark designed to identify and address safety vulnerabilities in long-context LLMs.

• Reveals that safety guardrails weaken as context length increases
• Demonstrates how adversarial content hidden in long contexts can evade safety mechanisms
• Proposes specialized alignment techniques for extending safety to long-context scenarios
• Shows that current LLMs (including GPT-4) remain vulnerable to safety attacks in extended contexts

This work addresses a critical gap in AI security by highlighting how safety challenges differ in long-context environments, providing essential guidance for securing next-generation AI systems that process extensive documents or conversations.

Original Paper: LongSafety: Enhance Safety for Long-Context LLMs