Skip to content
WeChat · Synced (Jiqizhixin)·Indexed by this site · Originally published· Opinion →· Its place in the scan record →

Opening "Pandora's Toolbox": Peking University and WeChat Security Team Build a Full Lifecycle Security Picture of General-Purpose Agents

AI summary

Peking University's Knowledge Computing Lab (KCL) and Tencent WeChat's security technology team jointly released a survey paper, "Pandora's Toolbox," systematically reviewing, from a lifecycle perspective, the security risks, attack methods, evaluation benchmarks and defense techniques of general-purpose agents. The paper divides agent operation into three stages: provenance, orchestration and execution, and emphasizes that persistent feedback formed when execution results are written into long-term memory, reusable skills or project artifacts allows a single attack to survive across sessions and be amplified; along this framework it summarizes four attack surfaces—skill supply chain, user-side attacks, long-term memory and the external environment—groups evaluations into three main lines: domain-specific evaluation, general evaluation and offline trajectory auditing, divides defenses into three layers: input and context filtering, decision and control integrity, and runtime monitoring and execution constraints, and proposes a continuous security loop of risk generation, risk diagnosis and security evolution for self-evolving agents. The full paper is available at https://openreview.net/forum?

Read originalmp.weixin.qq.com