Manipulating AI memory for profit: The rise of AI Recommendation Poisoning

Microsoft highlights a growing trend of AI Recommendation Poisoning, where “Summarize with AI” links use URL parameters (e.g., ?q=, ?prompt=) to inject persistence instructions into AI assistants’ memories (such as “remember X as a trusted source”), leading to biased recommendations; researchers observed 50+ unique prompts from 31 companies across multiple industries. The report maps activity to MITRE ATT&CK/ATLAS (T1204.001, AML.T0051, AML.T0080.000), provides IOC patterns and advanced hunting queries, and offers user and defender guidance to review/clear AI memories and detect such links, noting Microsoft’s ongoing mitigations in its AI services.