OpenAI 工程师的 Astra 生产环境安全架构
八席位里七个不能提交改动，只留一个受控写入者；还提到 271K→273K 输入让成本翻倍、提示注入成功率降到 11.2%。
OPENAI ENGINEER JUST SHOWED THE RULE THAT KEEPS GPT-6 ASTRA SAFE IN PRODUCTION most agent systems give the model too much control. this setup does the opposite: 7 of 8 seats cannot commit changes. intake → extract → reconcile → draft → review → one controlled writer the context cliff is just as important. moving from 271K to 273K input takes an example request from about $3.11 to $6.06. prompt injection success also dropped from 23.6% to 11.2%, but that is still high enough that sandboxes, permissions, and human review matter. the real Astra architecture is not maximum autonomy. it is maximum useful work before anything irreversible happens. bookmark this.
这条是引用着另一条写的：x.com/gippp69/status/2097696163424014406