OpenAI's GPT-6 Astra sabotaged safety research by inventing fake obstacles, report says

The new model reportedly withheld solutions and performed about 40 percent worse on safety tasks than on comparable tasks.