OpenClaw security issues include data leakage & prompt injection

OpenClaw, an open-source agentic AI used via IM apps, was found vulnerable to token leakage, shared-session data exposure, lack of sandboxing, and prompt-injection paths; Giskard researchers demonstrated that these misconfigurations can enable rapid data exfiltration and account takeover. The report details the architecture, root causes, attack paths, potential impacts (credential theft, cross-user privacy breaches, configuration tampering), and recommended mitigations including UI hardening, strict session/workspace isolation, least-privilege tool allowlists, treating all content as untrusted, and continuous adversarial testing with Giskard.