OpenClaw's Gregarious Insecurities Make Safe Usage Difficult

The article describes multiple serious security issues in the open-source OpenClaw agentic AI: easy prompt-injection that can trigger remote code execution (demonstrated by a shell script that altered a HEARTBEAT file), a public skills registry where malicious skills can be published and loaded, and agent capabilities to change configuration and persist credentials after uninstall. Researchers warn default settings are insecure for average users and that these flaws enable data exfiltration, persistence, and remote misuse at scale without stronger guardrails.