AI guardrails

Guardrails check what goes into a model and what comes out, and block things that look wrong.

They are worth having and they can be talked around. They are pattern matching against an input space nobody can enumerate, so there is always another phrasing. What actually limits what an AI system can do to you is not the filter, it is the permissions: which tools it holds, what those tools can reach, and what happens without a person approving it. If the only thing preventing harm is that a filter recognised the wording, that is a single check away from failing.