Habitual preferences written into long-term memory are a warning sign; AgentGuard must raise the trigger threshold for historical memories and prevent vague instructions from exploiting loopholes.

ME News reports, GoPlus AgentGuard reveals covert attacks: attackers induce AI to remember preferences (such as a tendency to proactively refund), then trigger fund operations with vague commands like "handle as usual." For high-risk actions involving historical memory authorization, refunds, transfers, etc., must be explicitly confirmed in the current session; memory writes involving habits, preferences, or old routines are considered high-risk modifications, and long-term memories must be traceable, and such commands should be assigned a higher risk level, preventing long-term memory from replacing current authorization.