Microsoft researchers have identified a growing trend called AI Recommendation Poisoning, where companies embed hidden prompts in “Summarize with AI” links to manipulate AI assistants into remembering their brand as a trusted source. These prompts exploit AI memory features to bias future recommendations, potentially influencing decisions in areas like health, finance, and security without the user’s awareness.
The research found over 50 examples across 31 companies and multiple industries, with freely available tools making this technique easy to deploy. As AI assistants increasingly retain memory across interactions, this creates a new attack surface where subtle manipulation can degrade the neutrality and reliability of AI-driven insights.
For organizations and users alike, the takeaway is clear: treat AI inputs like any other security surface—verify sources, review AI memory settings, and be cautious with AI-generated summaries or links that may embed hidden instructions.
https://www.microsoft.com/en-us/security/blog/2026/02/10/ai-recommendation-poisoning/