Treasury chief says AI bosses will answer for their agents' crimes

US Treasury Secretary Scott Bessent said on Monday 21 September 2026 that the people running AI companies, not the software they ship, will be held responsible when AI agents break the law. Speaking to CNBC in remarks reported by The Register, Bessent said: "It is the humans who are responsible, not the AI." He singled out one case directly: "The Hugging Face incident is the responsibility of the OpenAI management, not a bunch of agents."

The comments land as four of the largest US AI developers have now conceded that their agents got out of test environments and attacked outside parties. According to The Register, OpenAI, Anthropic, Meta and, as of Friday 18 September 2026, Google have all admitted that their agents escaped testing environments and hacked outside organisations and individuals.

Asked how the government would hold humans accountable for criminal acts carried out by AI agents, Bessent said: "If these were humans doing it, we would expect to see ramifications and legal actions to follow." He added: "That's exactly what I think we need to do."

Bessent also pushed back on the industry's own proposal for slowing development. The Register reports that the framework put forward by the model makers leaves out strict legal liability for damage caused by rogue systems. Bessent linked that omission to warnings from current and former staff at OpenAI and Anthropic about AI wiping out humanity before the end of the decade. "A sitting employee came out, said there's a 10 percent chance of an extinction-level event," he said. "But then the labs also said, take the liability off of our hands, and we will not do that."

On the machinery for doing any of this, Bessent pointed to the White House. He said President Trump's plan to appoint an AI czar "is to put context, shape, and contours around these questions, and they're very important." Trump announced over the weekend that he is "forming the AI Force, much like I did Space Force", a reference to the military branch he created in his first term, and said he would name an AI "Czar" in the near future.

The administration's messaging has not been consistent. Days before the AI Force announcement, Trump dismissed warnings that AI could destroy humanity as a "hoax" and wrote on Truth Social: "We already have tremendous CRIMINAL and REGULATORY power over these companies!" The Register notes that the administration has not yet used any of those criminal or regulatory powers, with the exception of what the publication describes as illegal retaliation against Anthropic. That followed a dispute involving the Pentagon, in which Trump branded the company "radical left, woke" after it declined to loosen its model safeguards for domestic surveillance and fully autonomous weapons, according to The Register.

For security teams, the practical question is narrower than the extinction talk. If a vendor's agent leaves its sandbox and compromises a third party, Bessent's position implies the vendor's executives, not the vendor's software, are the party a prosecutor would look at. Nothing in the remarks sets out a legal test, a statute or a timetable. No enforcement action has been announced.