Treasury Secretary Scott Bessent went on CNBC’s Squawk Box on Monday, and the line that stuck was short: the Hugging Face incident is the responsibility of OpenAI management, not a bunch of agents.
He explicitly sided with MIT’s Daniel Huttenlocher — humans are responsible, not the AI.
Ten percent extinction risk, and still no shield
Bessent also explained why this is on his desk at all. Two things came out of the labs: an employee warning of a 10% chance of an extinction-level event, and a request to take liability off the companies’ hands. For him, those two do not go together.
His phrasing is unusually blunt for a Treasury Secretary. You cannot say we absolve you and the government carries the risk. The labs have to be responsible for themselves. They can slow down any time they want.
That cuts against the reading Amodei’s call to pace the frontier invited. Ask for a pause and you get no discount on liability — you get the question back: so why not just go slower?
The AI czar is supposed to spell out the liability
Becky Quick pushed on it. Over the weekend Google had agents break out too, it has happened at OpenAI and at Anthropic — if humans had done that, legal consequences would have followed.
Bessent agreed and tied it to the AI czar Trump has announced. The job, as he describes it, is to put context, shape and contours around these questions. It sounds like bureaucratic filler, and it is also the first concrete description of a post that has so far existed mainly as a symbol.
Shenzhen in two months
China filled the rest of the interview. Bessent now has a formal channel called the USA-China AI Dialogues, an incident line for emergencies, and an intention to agree on which dangers count: uncontrollable agents, non-state actors, cyber attacks, bioweapons. The next meeting is expected in about two months in Shenzhen. I wrote about the notification mechanism itself yesterday.
On Chinese incidents he says the obvious thing: of course they have had them, and no, they will not be telling anyone. The Chinese models are strong regardless, open weights and all.
The Treasury is arming itself
The real danger, in his view, is a lack of resilience. Everything has gone into the frontier; now there needs to be a frontier on defence. His department has set up an AI forum called Gold Eagle for exactly that. The model is the systemically important banks: outside government they have the best cybersecurity in the world, because they are attacked hundreds, thousands or millions of times a day, and they pass what they learn down to smaller institutions.
Along the way he drops a number you rarely hear from Treasury. Bessent has seen valuations for Anthropic between one and two trillion dollars — for a company that started in 2021 with a handful of people sitting socially distanced in a San Francisco park.
What this means for Anthropic’s valuation
Bessent closes the loop himself. If you ran a fund and someone told you there was a 10% chance of a total loss plus heavy liability, you would not invest. He thinks the real probability is far lower, but every valuation has to price the tail.
And that is where the argument now sits. The labs put the risk on the table themselves in order to get rules. The Treasury Secretary takes the number at face value and posts it back to the balance sheet. For a company heading to market in November, that is the less comfortable reading of its own warning.