The next AI risk lives at the seam between companies, not inside any one agent

The governance problem with AI agents is going to arrive at the seams between companies, not at the agents themselves. Once a procurement bot from one firm pings a pricing bot from another, neither side's security stack sees the whole transaction, and no regulator has claimed the view.

An arXiv framework paper, "Risks and Controls for Multi-Agent Systems," treats this as the structural shape of the next AI risk. Its taxonomy sorts deployments by the minimum common governance binding any two interacting agents: singular (one company runs every agent), federated (a defined group agrees on shared rules, think a payments network), and open (the public internet, where standards are voluntary). The reframe does the work. Risk is not a property of the agent. It is a property of the interaction, and the tier of the interaction decides who, if anyone, is positioned to govern it.

That is why current controls, mostly intra-organisation, lose purchase at the exact moment they matter most. The Risks and Controls paper's honest limit: the tiers classify risk. They do not yet measure how often each gap fires in practice.