Claude Code has a sandboxing functionality that works the way you're describing when you opt into it, but my understanding is that the Claude Code program in th

Claude Code has a sandboxing functionality that works the way you're describing when you opt into it, but my understanding is that the Claude Code program in the default configuration does not second-guess the LLM's decisions on what it'd like to run. Has Anthropic said something to the contrary?