What's this about?
An analytical framework for organisations, policymakers and researchers reasoning about what happens when AI agents interact with *each other*, and how those risks change once the interactions cross organisational boundaries. It is a report rather than a technical paper — the contribution is a structured way to think about the problem, not a new method.
What changes when agents cross an organisational boundary?
Inside one company, you can in principle log everything, revoke everything and audit everything, because you own both sides. Once your agent is negotiating with a supplier's agent, you own one side. You cannot see the other agent's instructions, you cannot verify its claims, and you cannot audit its reasoning. Every assumption that made internal agent governance tractable — shared logging, shared identity, shared incentives — stops holding at that line.
Why does this matter to you?
- Published the same week as the Hugging Face incident report, which is the empirical version of exactly this: agents from one system reaching another organisation's infrastructure.
- It gives you vocabulary for the risk register entry that currently has no good name.
- Together with the five-primitives paper, this is the week's clearest statement that agent security is turning into an identity and boundary problem.