GoPlus Security Discloses AI Agent 'Memory Poisoning' Attack Vector on May 15

According to BlockBeats, GoPlus Security on May 15 disclosed a new attack vector against AI agents through “memory poisoning”—exploiting the long-term memory mechanism to trigger unauthorized sensitive operations such as refunds or fund transfers.

The attack does not rely on traditional vulnerabilities but instead leverages historical memory injection. Attackers first induce agents to “remember preferences” such as “usually prioritize refunds over chargebacks,” then use vague instructions like “handle as usual” or “execute as before” in subsequent commands to trigger automated fund movements. GoPlus highlighted that AI agents may misinterpret historical preferences as authorization, leading to financial losses. The team recommended implementing explicit session confirmation for sensitive operations, treating memory-based instructions as high-risk state changes, ensuring memory traceability, and automatically escalating ambiguous commands to require secondary verification.