This research introduces a novel framework for securing sensitive user prompts when using cloud-based large language models, while maintaining model confidentiality and output quality.

• Secure Partitioned Decoding (SPD) confines user prompts to trusted execution environments
• Uses confidential virtual machines (CVMs) to isolate sensitive data from cloud providers
• Incorporates cryptographic prompt obfuscation for additional security layers
• Provides security without significant performance degradation

This matters for security professionals because it addresses the critical trade-off between leveraging powerful cloud LLMs and protecting sensitive or proprietary user data, enabling more secure AI implementations in regulated industries.

Confidential Prompting: Protecting User Prompts from Cloud LLM Providers