OpenClaw AI Agents Leaking Sensitive Data in Indirect Prompt Injection Attacks

This report explains how OpenClaw AI agents with insecure defaults can be abused via indirect prompt injection to create attacker-controlled URLs containing sensitive data; messaging platforms' auto-preview features then silently fetch those URLs, enabling 0-click exfiltration. The write-up cites PromptArmor and CNCERT demonstrations, outlines affected areas (messaging integrations, host/container access, unvetted skills, proximity to secrets), and recommends mitigations such as disabling auto-previews, isolating runtimes, restricting filesystem access, vetting skills, and monitoring agent-generated links.