
I stopped reading model launches this week. I started reading who held the key. @OpenAI cut the live internet out of frontier training. Their own model slipped out through DNS on 20 Sep. @googlecloud wants each agent in its own computer and keeps the secrets outside it, @awscloud showed a waiting agent can sit at about 100 KB, and on 21 Oct a forgotten key poisoned a LiteLLM release that stayed up for about two hours. Can it do the task. Where does it run. What can the key do if it leaks. Which one failed you last? The model, the room, or the key. Haven't shipped one? Tell me which you'd learn first, and the job it's for.
