I completely understand the need for guardrails, but in practice they can significantly reduce the tool's usefulness for legitimate scientific research. It sometimes feels like the model is better suited for "vibe coding" than for serving as a serious research assistant.

What's more funny is that you can use claude code together with molecular design models like RFdiffusion and Alphafold to develop a real bioweapon, but since this is in the realm of software, the prompts do not get flagged at all. Those models can even run on consumer GPUs so the limit of access is almost not there.

Has anyone else had a similar experience?