OpenAI's automated research pipeline used an upcoming cyber model to identify 1,724 flaws and produce 948 working root exploits on the Linux kernel