prompt poisoning

Prompt poisoning is a security vulnerability where hidden instructions are injected into AI assistants through prefilled prompts, often via features like 'Summarize with AI' or 'Share with AI' buttons. Microsoft's security researchers have documented how real companies exploit these mechanisms to bias AI memory and recommendations, making malicious guidance appear neutral. This tag covers discussions on the mechanics of prompt poisoning, its detection, and mitigation strategies within enterprise and consumer AI tools, with a focus on Windows and Microsoft ecosystems. Topics include AI security, memory manipulation, and the challenges of distinguishing poisoned outputs from legitimate AI responses.