
OpenClaw agent hands its secrets to anyone who asks nicely in the group chat
Ourtake
Privacy was one setting away, and the default was 'main'.
What happened
Researchers at Giskard tested a deployment of OpenClaw, an open-source agentic AI assistant reachable through chat apps. They found that default session sharing, weak tool sandboxing and prompt injection through external content let secrets and other users' data leak. The article presents these as design and configuration failures, with mitigation guidance.