Security researchers found more than 21,000 exposed OpenClaw deployments, highlighting why enterprise adoption of AI agents needs to pay more than lip service to "guardrails".