They fall into the categories of "hard" or "soft", and if you use AI at all, you should be very clear about which guardrails apply.
"Hard" guardrails involve withholding capabilities. That could mean system permissions or sandboxing. That means they are imposed from outside the AI, and they tell it whether it's allowed to save a file, delete an email, or run a program.
"Soft" guardrails are from within the AI environment. Prompts like "Don't talk about Bruno" provide a lens through which the AI can answer, but it doesn't actually stop it from discussing Bruno. A soft guardrail may not be effective if the constraint is not dominant relative to the rest of your query.
The difference between hard and soft guardrails is the difference between handing someone a phone and saying "If this rings, don't answer it" and not giving them a phone in the first place.
So the question you need to ask when someone tells you there's a guardrail is "What happens if the model ignores the guardrail?"
If the answer is everything other than "it literally can't" it's a soft guardrail, and you should not rely on it as a 100% guarantee.
Soft guardrails are still useful as they can, in general, steer the model to do what you want. But if you need to actually control the AI's behavior then hard guardrails are absolutely
necessary.