OpenAI stated that the GPT-6 model showed stronger resistance to attempts to bypass its safety training compared to previous models, including its predecessor GPT-5.6