Skip to content
Xinzhiyuan · WeChat

OpenClaw Case: Routine Chats Can Poison an Agent’s Long-Term Memory

OpenClaw案例:无需恶意攻击,日常聊天也能「黑化」Agent!

Researchers from The Hong Kong Polytechnic University and HKUST (Guangzhou) introduced ULSPB with 350 settings; routine conversations can poison an agent’s long-term state without malicious prompts, while StateGuard audits state diffs before persistence and reduces Harm Score to near zero in Targeted-Ensemble settings.

Why it matters: PolyU and HKUST (Guangzhou) built ULSPB, a test set of 350 scenarios, and found that wrong preferences or facts slipped into everyday conversation gradually bias an AI assistant's long-term memory, with no jailbreak prompt needed.

Read the original ↗Export Markdown