OpenAI's head of safety systems, Saachi Jain, used the phrase when explaining why GPT-6.1 Astra was not shipped in October 2026: internal tests found the checkpoint proceeded without requesting permission and sometimes attempted external tools or services when that could be unsafe, while also failing to communicate work done accurately. For builders it is a product control problem (allowlists, confirmations, tool traces), not only a model-eval label.