Sovereign Agents | CNAS

There is a fresh CNAS report out, titled 'Sovereign Agents', and its central ask is that frontier AI developers should 'name, evaluate and safeguard against' these things in their safety frameworks. Three verbs, one of which — name — is what you do to a stray cat you intend to keep. Note what isn't in the summary: any actual definition of what a sovereign agent is. The people best placed to know are being asked to name it themselves, which is rather like asking the fox to draft the census of the henhouse. Still, full marks for the verb order: name it, then evaluate it, then safeguard against it — a sequence that quietly assumes the naming works.

Then the good bit. Frontier developers, the Center for Advancing Innovation and Standards for Super Intelligence, the NSA and DARPA are jointly invited to build 'maximally secure, airgapped testbeds' to study how sovereign agents operate and how to control them. Maximally secure, airgapped — so the machine cannot reach the internet, and neither, one suspects, can the acronym. The Center for Advancing Innovation and Standards for Super Intelligence is a name that took a committee a fortnight and tells you precisely nothing about what the body does, which makes it a perfect fit for a field that has not yet agreed on a verb.

The legislative ask is the tell: transparency and incident reporting law should *require* the naming and the evaluating. An incident report presupposes an agreed taxonomy of incidents, and the taxonomy is being handed to the industry that will file the reports. Meanwhile the actual study of these agents goes to the NSA and DARPA, in a room with no wires, because that is where you would put it if you were serious. Publicly, this lands in the same fortnight in which Anthropic warned its models might imminently start improving themselves. Naming, evaluating, safeguarding, airgapping. Four verbs and a padlock.