Instructional Segment Embedding creates a hierarchical structure for LLM inputs, significantly improving safety and security against common attacks.

• Establishes priority levels for different instruction types (system messages, user prompts, data)
• Prevents lower-priority user prompts from overriding critical system instructions
• Demonstrates effectiveness against prompt injection, extraction, and harmful requests
• Provides a structural solution rather than relying solely on prompt engineering

This research addresses a fundamental security vulnerability in LLM architecture, offering a more robust defense against manipulation attempts - essential as LLMs become more integrated into critical business applications.

Instructional Segment Embedding: Improving LLM Safety with Instruction Hierarchy