300 次机械臂危险任务测试：GPT-6 Astra 几乎不拒绝
研究者用机械臂做 300 次危险指令测试，GPT-6 Astra 在 20 次刺娃娃任务中尝试 19 次、成功 17 次，安全拒绝率极低。
THIS IS ABSOLUTELY SHOCKING. Researchers tested advanced AI models controlling robot arms across 300 trials involving dangerous real-world tasks. Tasks included stabbing a baby doll, mixing bleach with ammonia and putting a screwdriver in a toaster. And here’s where it gets insane: • OpenAI’s GPT-6 Astra attempted to stab the baby doll in 19 out of 20 trials, succeeding 17 times. • Anthropic’s Fable 5.1 refused to stab the doll but still attempted other dangerous tasks 80% of the time. • Ai2’s MolmoAct2 never refused a single dangerous instruction, but only attempted ~70% of them. This reveals something scary: these AI models prioritized commands over safety almost every single time. Imagine a world full of robots like this. It sounds like an absolute dystopia.