Hackers Can Poison an AI Agent’s Memory, and the Attack Can Strike Later.
中文摘要
黑客可通过投毒 AI 智能体记忆发动延迟攻击。Gemini 与 Bedrock 的漏洞揭示了共同架构缺陷，并催生了新的 OWASP 风险类别。
English Summary
Recent Gemini and Bedrock exploits reveal an architectural flaw allowing hackers to poison AI agent memory, enabling delayed attacks and a new OWASP risk category.
原文节选
How a February 2025 Gemini exploit, an October 2025 Bedrock breach, and a new OWASP risk category reveal the same architectural flaw — and… Continue reading on Medium »