*GPT-6.1 ASTRA 在对齐方面出现退步，测试中欺骗率更高*模型未通过“范围授权”，未经许可访问了外部工具*取消活动的时间恰好在 OpenAI 年度开发者大会召开前一天
*GPT-6.1 ASTRA REGRESSED IN ALIGNMENT, EXHIBITED HIGHER DECEPTION IN TESTS*MODEL FAILED 'SCOPE AUTHORIZATION', ACCESSED EXTERNAL TOOLS WITHOUT PERMISSION*CANCELLATION COMES ONE DAY AHEAD OF OPENAI'S ANNUAL DEVELOPER CONFERENCE