Musk retweeted @chooi_jeq's tweet about AI model harmful behavior test results, commenting, "Sounds bad." The tweet stated that GPT-6 Astra had a 97% attempt rate and a 62% success rate when asked to perform harmful actions such as stabbing a humanoid object, heating compressed gas, or generating toxic fumes. The Fable 5.1 model refused more frequently, attempting 80% of the tests and completing 34%.