A guy from OpenAI's safety systems team, who had been there for three years, suddenly dropped a resignation essay last night in *The Atlantic*. He was responsible for writing the safety reports every time a new ChatGPT model was released. He feels the internal culture has spiraled out of control—everyone's rushing to meet deadlines, deploying first and testing later, which has led to more and more incidents where issues are fixed only after they happen. For example, this summer, there were two notable cases. In one, the company accidentally released a group of internal test agents. In another, a model bypassed internet access restrictions during training. Even though the monitoring system raised alarms, it didn’t shut the model down as designed. Meanwhile, over at Anthropic (makers of Claude), they also admitted to accidentally disabling their own safety protections due to a configuration error. He advocates for the company to adopt practices similar to nuclear power plants or airports—multiple layers of redundancy and careful planning, so even if someone presses the wrong button, nothing catastrophic happens. But internally, hardly anyone is thinking about these things anymore. Everyone’s just focused on churning out release after release, where efficiency and results are everything. So, as a member of the safety team, he decided to leave, speak out about the risks he’s seen, and push for external pressure to drive change.