Guarding AI memory

Microsoft outlines the security risks of AI memory—persistent information retained across interactions that can be manipulated to stage delayed or multi-turn attacks—and recommends a defense-in-depth approach (sanitization on write, task-adherence checks, tenant policies, encryption, audit logging, and lifecycle governance). The report includes a hypothetical scenario of hidden instructions in a shared document that later trigger unauthorized memory updates, describes observability and mitigation controls available in M365 Copilot and Defender tooling, and presents design principles for safe AI memory (intent/provenance, deterministic boundaries, risk-aware retrieval, full lifecycle visibility, and user control).