The Trump administration spent the first half of 2026 telling the AI industry: build fast, we will not get in the way. Then Anthropic released Mythos, and everything changed.

Mythos, Anthropic's most powerful model, can spot decades-old security vulnerabilities that human auditors missed for years. It is, by virtually every measure, the most capable AI system ever released to the public. And it rattled the White House enough that administration officials began privately discussing something they had previously mocked: pre-release government vetting of AI models before they ship.

The problem? Nobody can agree on what that means, and the industry is panicking.

Seven Lobbyists, Zero Answers

Politico reported that seven lobbyists and policy advisers told the outlet they have struggled to get specifics on a possible executive order that could impose vetting rules on new AI models. The confusion is not about whether the White House is considering it. Multiple sources confirmed discussions are underway. The confusion is about what form it would take, who would run it, what models would be covered, and when it would apply.

"Lack of organization" is how Politico described the situation, and that is generous. This is the same administration that revoked Biden's AI safety executive order on day one, called AI regulation a threat to American competitiveness, and positioned itself as the anti-regulation alternative to Europe's AI Act. Now it is quietly exploring the exact kind of oversight it campaigned against.

Mythos Changed the Calculus

The trigger is clear. Anthropic's Mythos did something that no previous AI model had done in a way that was publicly visible: it demonstrated offensive security capabilities that made the national security establishment uncomfortable. Finding vulnerabilities that have existed for decades is useful in a defender's hands. In an adversary's hands, it is a weapon.

The administration's typical hands-off approach was built on the assumption that AI models are mostly chat tools and coding assistants. Mythos punctured that assumption. And the Pentagon's decision to blacklist Anthropic from defense contracts (which we covered in prior cycles) did not reduce the model's capabilities. It just meant the government could not see what was coming next.

The State-Level Stampede

While the White House dithers, states are moving. This week alone: Colorado passed SB 26-189, a comprehensive rewrite of its AI consumer protection law covering automated decision-making in employment, lending, insurance, and housing. Connecticut's SB5, which we covered last cycle, bans AI chatbots from romantic interactions with minors and mandates watermarking for synthetic media. Both bills passed with strong bipartisan margins.

The emerging picture is a regulatory vacuum at the federal level being filled by a patchwork of state laws. It is the worst possible outcome for the industry: instead of one set of rules, they will have 50. And the White House's inability to articulate a coherent position is why.

What Happens Next

The Nadella testimony in the Musk v. OpenAI trial, which began today, is going to add pressure. Congress is watching. Seven AI-focused bills are active at the federal level. The Senate AI working group has hearings scheduled for later this month with Ilya Sutskever on the witness list.

The irony is thick: Europe just watered down its AI Act after heavy lobbying from the same tech companies that are now begging the US government for clarity. They do not want regulation. But they want uncertainty even less. An executive order with clear rules, even restrictive ones, would be better for business planning than the current state of affairs, where nobody knows what is coming and the answer changes depending on which White House staffer you ask.

The administration that promised deregulation might end up delivering the one thing worse than regulation: chaos.

Source: Politico, WKRN (The Hill), Colorado General Assembly (SB 26-189), The AI Post prior coverage.