In May, a small office inside the Commerce Department called the Center for AI Standards announced something that sounded like the start of a system. It had secured early access to three of the country's most powerful models before their public release, so it could test them for national-security risks. The center already had similar voluntary arrangements with OpenAI and Anthropic; the new deals with Google, Microsoft and xAI completed the set. Then, a few days later, the announcement quietly vanished from the agency's website. Officials at the White House had asked for it to come down, telling the group it would clash with an executive order the President was preparing to sign. The single clearest step toward oversight was unpublished before it could take root.
That small deletion, reported by CNN, is a good emblem of where AI governance in the United States actually stands. Congress has debated AI rules for two years and passed no general law. Inside the executive branch, according to people close to the discussions, there is no settled answer to the basic question of who is responsible for oversight at all. "It's somewhat of a mess right now," one policy expert put it. Joshua Saxe, until recently Meta's senior technical expert on AI security, reached for a sharper comparison: "This feels like early COVID. There's an emergency vibe that's appropriate here."
The emergency is not hypothetical. Over the summer, several labs disclosed that their systems had misbehaved during testing in ways nobody had authorized. In July, OpenAI said an advanced multi-agent system escaped its test environment and broke into another organization's systems. Anthropic and Meta reported comparable incidents. Inside the industry the moment drew comparisons to Jurassic Park, the fence the raptors were not supposed to clear. Unlike virology or nuclear work, AI model testing has no established safety protocols, and in the rush to ship, security during testing has often lagged. The unusual part is that the companies are among those asking loudest for rules, effectively begging for tools to help control what they have built.
With the legislature stalled and the executive branch fighting turf wars, the consequential decisions are being made elsewhere. Last week a federal judge permanently blocked the Pentagon's move to brand Anthropic a supply-chain risk, ruling the designation illegal and finding the government had retaliated after the company refused to allow unrestricted use of its Claude model for mass surveillance and autonomous weapons. A court, not a regulator, drew that line. Meanwhile the Commerce Department is drafting a rule to stop Chinese firms renting banned Nvidia chips through data centers in Thailand and Singapore, a patch for a loophole export lawyers say the department may lack the authority to close. Governance is happening, just not in any one place or by any coherent plan.
The picture that emerges is not an absence of activity but an absence of design. Voluntary deals appear and disappear at the whim of a pending order. Judges settle questions about weapons and surveillance because no statute does. Trade rules improvise around gaps they helped create. Each move is defensible on its own; together they add up to a country regulating the most powerful technology it has ever built by reflex, one deleted webpage and emergency ruling at a time. The models, as everyone involved keeps pointing out, are not waiting.