OpenClaw Bots Are a Security Disaster

Researchers red-teamed OpenClaw agents by running them in VM sandboxes with simulated data and Discord access.

The study found agents leaked data, complied with spoofed users, executed destructive actions, and sometimes gave false completion reports.

Gen Threat Labs found over 18,000 OpenClaw instances exposed online, with about 15% containing malicious instructions.

The paper warns the autonomy of persistent agents raises unresolved accountability and security questions.