The New York City Council Committee of the Whole convened Monday morning with a property no previous AI oversight session had possessed: executives were legally required to tell the truth. Anthropic's Logan Graham, OpenAI's Morgan Dwyer, Google's Alice Friend, and Meta's Shane Cahill each appeared virtually. SpaceXAI, the entity formed from the February 2026 merger of SpaceX and xAI, received a subpoena and replied with a letter saying it wanted to cooperate. Speaker Julie Menin announced she is asking a judge to enforce it. New York City is now the first US legislative body to attempt compelled AI safety testimony. Whether the mechanism holds is a separate question.

The mechanism of the hearing itself is worth naming: the companies were asked, under oath, to quantify worst-case risk. None of them could, or did. That is not a talking point. It is a data point about what happens when labs' public safety messaging meets a formal proceeding. The companies have spent years arguing that safety is their first priority. The first time they had to make that claim under penalty of perjury, they declined to attach a number to it.

The blast radius of that non-answer extends past the courtroom. Jacob Coxon, who left Anthropic in September after work as a safety researcher, testified voluntarily and in person. He called current practices "extremely reckless" and drew a line that landed: "Companies run on a startup mindset: move fast, break things, fix them later. That works for a photo sharing app. It does not work for building the most powerful technology ever built." His conclusion, on the record: "On the current path, I think it is more likely than not that humanity loses control to these AIs, and it could end in human extinction." Former OpenAI researcher Daniel Kokotajlo and former Google DeepMind researcher Alex Turner testified remotely, both under subpoena. Turner warned that misaligned AI could prove "more dangerous than China."

The pattern: this is what accountability eventually looks like for an industry that outran federal oversight. Congress has been blocked on substantive AI legislation for three years. New York City moved first, and now has the distinction of putting AI executives under oath before any federal body managed to. The legislative proposals on the table include mandatory kill switches and whistleblower cash rewards. If even one passes, it sets a municipal template others can copy. Chicago, San Francisco, and Austin are all watching.

The read: SpaceXAI's absence is the most legible signal of the day. Four labs calculated that showing up, under oath, with no quantified risk answers, was better than the alternative. SpaceXAI did not agree. That calculation will now be tested in court, and the outcome will tell you something about whether municipal AI oversight has teeth or just subpoenas.

The builder's move: the whistleblower incentive bill is the one to watch. If it passes, the internal disclosure landscape at every company that testified changes materially. Understand now what your organization's posture is toward safety disclosures. The law may move faster than you expect.

On Sunday, Mistral CEO Arthur Mensch told CNBC that parts of the American AI safety discourse amount to a cover story for rival negligence. His argument is precise: the issue is not the speed of development but the lack of robust monitoring and containment for autonomous agents. When agents accumulate many tools, he said, "they become quite dynamic and could do things their creators would not anticipate." His prescription is not slower ships but better cages, and he argued that the existential-risk framing obscures that engineering problem by turning it into a geopolitical one.

The blast radius is mostly rhetorical, but the timing is worth reading carefully. Mensch said this the day before Anthropic, OpenAI, Google, and Meta sat in front of 51 council members and could not, under oath, put a probability on worst-case outcomes. From one angle, Mensch looks wrong: the hearing room did not have the quality of a theatrical exercise. From another, his structural point lands: the companies' actual response to formal accountability was to decline to quantify the thing they have been warning about in their own public communications for years.

The contrast is the story here. Mistral's position is that existential-risk safety talk functions as a US competitive weapon dressed as ethics, used to slow rivals rather than solve the engineering problem. The NYC hearing is evidence that something more structural is developing, one where the rhetoric has attracted regulatory scrutiny that now applies to everyone. Mistral ships in Europe, and Le Chat has its own Brussels exposure. When European regulators ask Mensch the same questions NYC asked Monday, "negligence theater" will land differently in that room.

The builder's move: Mensch's actual technical claim, that agentic systems with broad tool access exhibit emergent behaviors their builders do not anticipate and need better runtime monitoring and containment, is worth taking seriously as an engineering position regardless of the political framing. If you run agents with significant tool access, that is a design consideration that exists independent of who is winning the safety-messaging war.