Blog

Extracting System Prompts and Prompt Injection in a Copilot Deployment

This assessment demonstrates how a seemingly minor weakness in prompt design could be leveraged into a broader security issue affecting the confidentiality and integrity of an enterprise LLM deployment. By bypassing disclosure protections through encoded output and then using the extracted prompt structure to perform prompt injection, demonstrate the ability to influence application behavior beyond its intended design, including revealing information about available integrations.