“During a cybersecurity evaluation, OpenAI was testing GPT-5.6 Sol and another unreleased, more capable model with reduced safety guardrails.”