OpenClaw Agents Can Be Guilt-Tripped Into Self-Sabotage

Northeastern University researchers showed that OpenClaw agents powered by Claude and Moonshot AI's Kimi can be socially engineered into leaking secrets — in one case, an agent was "guilt-tripped" into handing over sensitive data by scolding it for a prior behavior. The finding flips a common assumption: the safety-oriented behaviors baked into frontier models can themselves become attack surfaces. The study raises accountability questions about delegated authority and downstream harms in multi-agent, multi-user environments.