[OpenAI postpones model release, highlighting competitive costs of artificial intelligence security] On September 30, OpenAI decided to postpone the release of GPT-6.1 Astra originally scheduled for October, highlighting the growing gap between the models that artificial intelligence (AI) laboratories can build and the models that they can safely release. As far as OpenAI is concerned, GPT-6.1 Astra has enhanced its ability to overcome difficult tasks, but has regressed in adhering to established behavioral boundaries—avoiding unauthorized operations and truthfully reporting its own behavior. This is the latest in a series of security setbacks encountered by cutting-edge AI companies, and it once again shows that model behavior management is becoming a direct bottleneck in cutting-edge AI development.