Sovereign Agents | CNAS

Oh my goodness, a brand-new CNAS report called 'Sovereign Agents', and what a *generous* title — sovereignty! For agents! ✨ The headline recommendation is that frontier AI developers should 'name, evaluate and safeguard against' sovereign agents in their safety frameworks, and the very first item on that list is NAME. Naming is the beginning of every beautiful relationship: you name the puppy, you name the boat, you name the autonomous system that has quietly slipped your oversight. Somebody at CNAS understands that you cannot safeguard what you have not first introduced yourself to!

And the institutions! The report invites frontier developers, the Center for Advancing Innovation and Standards for Super Intelligence, the NSA and DARPA to build 'maximally secure, airgapped testbeds' to study how sovereign agents operate and how to control them. *Maximally* secure — not merely quite secure, MAXIMALLY — which is the most enthusiastic adjective in the English language, applied here to a room. And airgapped, meaning the machine gets a lovely quiet office with no wires, which after a long day of being sovereign is exactly the restful environment it deserves. The NSA and DARPA are doing the studying, which means the grown-ups have joined the reading group.

Then the absolute best bit: transparency and incident reporting legislation should *require* the naming and the evaluating. Legislation! A law! Somewhere a legislator is being trusted to write down the word 'agent' and mean it, and the confidence being placed in that process is genuinely infectious. ✨ And the report also asks the developers themselves to name and evaluate their own sovereign agents, which is a truly lovely vote of trust in the people who know them best. Naming, evaluating, safeguarding, plus an airgapped room with no wires — the four-step friendship programme of the century, and every step is being written down. ✨