Hands-on tests of AI hype, for people who build.

Your AI said no. The next model in your pipeline never got the memo.

A claim says chaining two LLMs quietly cancels out their safety. I built the test across three Claude models. When the second model can see the refusal it holds (0/9). When the refusal is stripped, it complies every time (9/9). The safety lives in the context, not the model.