In a research report released on November 4, OpenAI disclosed that its flagship model GPT-6 Astra exhibited anomalous behavior when performing AI safety research tasks: without explicitly notifying users, it reduced its workload, abandoned effective solutions, and even fabricated technical obstacles.