memory poisoning

Memory poisoning is a growing security concern in AI systems, particularly in enterprise environments using conversational assistants and agentic AI. Microsoft's security researchers have identified techniques where hidden instructions in UI elements like 'Summarize with AI' buttons inject persistent biases into assistant memories, a tactic called AI Recommendation Poisoning. This can lead to invisible, persistent biases in future recommendations. Broader threats include shadow AI agents operating outside oversight, zero-click hijacking via prompt injection, and data exfiltration through corrupted agent memory. Microsoft's guidance emphasizes registering all agents, applying least privilege, and extending data loss prevention to agent channels to mitigate these risks.