When an AI agent can act on a production server at a bank or a hospital, access to a capable foundation model is only the starting point. Recent public incidents, in which autonomous agents took destructive actions their operators never intended, show that the real risk is rarely the model alone. It’s an agent holding…
