One advantage is that this type of safety guardrail can't be undone by an adversary in post-training, so it's a good fit for open source models.
The experiments are all done in preventing models from acquiring medical capabilities, while preserving related capabilities like e.g., biology.