They aren't. They are saying that putting this safeguard at the level of the agent execution is a waste of time, particularly that of the researchers' expensive manhourse, especially now that existing smaller models can just evaluate the output.

It's fine if they seek to avoid that. But those are perfect examples of things a cheap output-filtering model can be responsible for enforcing, without wasting the time of the researchers who are working on the training and alignment of the model in the ways that actually matter.

Most "bad outputs" in areas like 'self harm' or 'violence' or 'sexual' whatever are a result of people deliberately trying to elicit those responses. As such, it's meaningless whether an LLM writes the Bad Thing or if the user writes it himself.