[_alternator_]

If data exfiltration is a danger in your threat model, you need local LLMs (or at least ones you fully control) not just the full chain-of-thought reasoning.