The Knowledge Computing Lab of Peking University, in collaboration with the Tencent WeChat Security Technology Team, has jointly released a comprehensive survey on general-purpose agent security. Moving beyond the conventional approach of categorizing security issues by component, they propose a unified, full-lifecycle analysis framework spanning provenance, orchestration, execution, and persistent feedback. This framework organizes four interconnected categories of attack surfaces: the skill supply chain, the user side, long-term memory, and the external environment. It distills the central thread of full-process security evaluation, proposes a three-tier defense system comprising input filtering, decision protection, and runtime constraints, and—to address the risk amplification challenge posed by self-evolving agents—constructs a dynamic security closed loop encompassing risk generation, diagnosis, and security evolution, thereby advancing agent security from single-point vulnerability patching toward full-lifecycle governance.
