Do those results reproduce with any other model than GPT-4o? A year-old paper analyzing a then-18-month-old LLM has dubious applicability to current LLMs, especially when some now have guardrails against persistently or arbitrarily rude users.