> Interleaved reasoning and function calling makes this even more dangerous. A model can call functions during the hidden reasoning phase.

The reasoning may be hidden but the tool calls are not, how else would the client execute them