Create an 'Agent-Trust' middleware that decouples the LLM's safety-check logic from the execution environment. This would allow users to define their own 'sandbox security profiles' to prevent over-zealous blocking by proprietary models.