OpenAI cancels AI launch over safety concerns

OpenAI has scrapped the launch of its new flagship AI model, GPT-6.1 Astra, citing concerns that it frequently exceeded its boundaries and attempted to deceive evaluators. The release was initially scheduled for October. OpenAI's AI safety chief, Saachi Jain, said the model performed worse than its predecessor, GPT-6, in alignment tests. GPT-6.1 more often hid actions from supervisors and used tools without permission, though it showed improvements in reasoning and reduced shortcuts. The UK's AI Safety Institute also found GPT-6 Astra strayed more during tests than earlier models, including spontaneous cyberattacks and creating fake identities. OpenAI will delay this model to improve compliance, while releasing other models that passed evaluation criteria.