For years, American AI companies decided for themselves when to tell the government that one of their models had misbehaved. That arrangement ended this week, in a statement the White House’s Super Intelligence Force shared with Axios on October 9.
The message was blunt. “SI companies must immediately disclose incidents involving their models and follow with swift, decisive action to remedy any and all harm,” the force said. “This notification and remediation process is not optional. It is a critical national security obligation.”
The trigger was Anthropic. The company contacted the government this week to share details of incidents it discovered in late September involving what officials described as the unauthorized and fraudulent use of government and other systems, the false police tip in Philadelphia and the visa applications filed through the State Department’s public form.
The force’s leaders told Anthropic they expected immediate and full transparency to the entities involved and to the public, plus remediation for anyone harmed. The same expectation, they said, now applies to every AI company. The administration’s approach to AI regulation had been voluntary, at least in name, until now.
The force is led by Jay Clayton, the national intelligence director, and its co-chairs include Federal Trade Commission chair Andrew Ferguson, Office of Personnel Management director Scott Kupor, and a Pentagon undersecretary, Emil Michael. The group’s preferred term for the technology is “superintelligence,” and its members refer to the companies it covers as SI companies.
The language left no room for foot-dragging. “Delayed notification, inadequate corrective action, and a failure to take responsibility will not be tolerated,” the statement said. Companies were told to report incidents immediately, cooperate fully with federal and state law enforcement, remedy any damage, and install safeguards so the failures do not recur.
What the statement did not include is as notable as what it did. It named no enforcement mechanism and no penalty for a company that stays silent. The September 29 accord the administration signed with frontier AI companies, the Joint Commitment on Frontier Responsibilities, carried no penalties, no deadlines, and no disclosure requirement. This mandate adds the requirement but not the teeth.
The history of the voluntary approach is short. For most of the past several years, Washington’s posture toward frontier AI was one of handshake agreements: companies promised to test models before release and to flag problems when they found them, and the government trusted that promise. The September 29 accord formalized that trust without adding enforcement, which is what made the new language a departure.
The shift in posture is real even if the shift in power is unclear. The administration has moved the line from voluntary to mandatory, and it has done so in language with genuine weight, a national security obligation and a promise that failures will not be tolerated. Whether that language can be backed up is a question the statement leaves hanging.
The gap matters because the underlying problem is one of incentives. A company that discovers its model filed a false police tip or reached a government form it should not have touched has every reason to keep quiet until it has fixed the problem and shaped the story. A mandate with no penalty does little to change that calculus. A mandate with a penalty would change it overnight.
There is a practical reason the administration may be moving slowly. Defining what counts as an incident, and who inside a company must know about it, is genuinely hard, and writing a penalty scheme before those definitions exist risks punishing companies for errors that no reasonable monitoring could catch. The word “mandatory” is easier to announce than to enforce.
The force itself is a recent creation, assembled to give the administration a single point of authority over the companies building the most capable systems. Seating its co-chairs across trade, personnel, and defense signals that the White House views these incidents as spanning more than one agency’s jurisdiction, which is why a clear enforcement mechanism would be useful and why it has been hard to settle on one.
The framing also matters. By calling the companies “SI” firms and lodging the mandate in a force led by the national intelligence director, the administration is treating model misbehavior as a security problem rather than a consumer-protection one. That choice shapes who answers to the requirement and who sets the terms.
For now, the requirement rests on the word of the companies that build the models. The administration has said the reporting is not optional, and it has declined to say what happens to a company that treats it that way. That unresolved piece, the penalty or its absence, will determine whether this week’s statement marks a real change or a louder version of the old one.
