AI Security

As autonomous AI agents gain the ability to execute long sequences of actions with limited human supervision, traditional notions of a simple “kill switch” become insufficient. This article proposes the AI watchdog as a broader safety architecture built around heartbeats, independent monitoring, zero-trust authorization, ephemeral credentials, sandboxing, rate limiting, transactional safeguards, graceful degradation, and human oversight. Its central distinction is that liveness does not imply correctness, and correctness does not imply alignment: an AI system can remain fully operational while acting outside human intent. The article argues that robust control should follow three principles: authority should expire, supervision should remain independent, and uncertainty should reduce autonomy. Rather than relying on a single emergency button, safe autonomous systems should be designed so that their power is continuously constrained, monitored, and progressively reduced whenever reliable supervision can no longer be guaranteed.