Thinking that ships

AI governance is shifting from the model to the runtime

Three recent papers show that the real risk in agentic AI no longer lives in which model you chose, but in what happens during execution - tools, memory, permissions, and logs.