LongSafety introduces the first comprehensive benchmark to evaluate and address safety vulnerabilities in large language models processing long contexts.

• Identifies 7 categories of safety concerns specific to long-context scenarios
• Reveals how safety risks can persist or emerge throughout extended interactions
• Demonstrates that current LLMs remain vulnerable to harmful content hidden within lengthy contexts
• Provides a foundation for improving safety mechanisms in next-generation long-context models

This research is critical for Security as organizations deploy LLMs with increasingly longer context windows, creating new attack vectors that existing safety measures may miss. LongSafety enables proactive identification and mitigation of these emerging risks.

LongSafety: Evaluating Long-Context Safety of Large Language Models