OpenAI’s latest GPT-5.6 safety results show low failure rates for direct prompt injection but higher success rates when attacks arrive through tools and external content. If you can only read one tech ...