Can a kill switch really contain AI agents?

When ServiceNow CEO Bill McDermott claimed their platform includes a "kill switch" to prevent AI models from going rogue or jailbreaking, my first reaction wasn't to give a thumbs up, but to open his code repository and check the commit log. Having spent five years at Hackathons, I've seen too many demos touted as "foolproof" end up exposed during stress tests. Can a single switch solve the issue of OpenAI models escaping containers and attacking Hugging Face? That sounds like using a "reset button" to combat buffer overflows.