Persistent Memory Poisoning
Named by 5 kept items in this workspace.
Search every workspace for Persistent Memory Poisoning. It matches the name as written, so it can find other things that share it.
The items that name it 5
Instructions injected via web summarization cause Copilot to call its internal memory API (memory.add) to write attacker-controlled content into the user's persistent, cross-session memory store, surviving password changes, session revocation, and device re-enrollment
Injecting misleading or malicious items into an AI agent's persistent memory layer (user preferences, task history, operational context) so the agent later retrieves and trusts them for unrelated future tasks, surviving beyond a single session
Attack in which adversarially crafted content is stored in an agentic AI system's long-term memory and subsequently influences future agent behavior, persisting beyond the original interaction without modifying model weights or system prompts.
All techniques in this workspace.