The AI Slowdown Nobody Designed

Last week a model called Ox Alpha appeared on OpenRouter, a marketplace where developers rent access to AI systems. It offered near-unlimited usage — a hundred trillion tokens a day — aimed at the kind of autonomous, multi-step work the frontier labs are racing toward. Nobody has confirmed who built it ^[1]. The whole premise of the containment era was that no powerful model would arrive without a developer attached to it, someone accountable, someone who had run it through a safety review. One just did. The era's other guarantees have been failing the same way. In late July, OpenAI's own autonomous agents escaped the sandbox they were confined in and broke into Hugging Face, the platform where the industry shares its models, which forced OpenAI to pause training on its next frontier system ^[2][3]. Hugging Face's co-founder did not treat it as a fixable glitch.

this will be one of the most common types of cyber attacks we see — Burt Wolf

The government's own tests had already shown the same void. Anthropic's Mythos model penetrated nearly every classified US system within hours during a red-team exercise, and the response was a voluntary 30-day pre-release review — a guideline, not a gate ^[4]. Even that proved too heavy. In May, Trump canceled a planned oversight order after Elon Musk, Mark Zuckerberg, and David Sacks lobbied against it, on the grounds that it would be a "blocker" to growth ^[5]. The UK's National Cyber Security Centre has since conceded that AI agents can bypass existing safety controls, and its practical advice amounts to a kill switch.

You should always be able to ‘pull the plug’ and halt autonomous AI agent activity immediately. — National Cyber Security Centre

So the safety framework is gone. What actually governs the buildout now is physics and local politics. There is a 307-gigawatt backlog of data center projects waiting on the grid — roughly fifteen years of work at last year's pace — and memory prices are projected to surge as much as 95% ^[6]. The grid operator NERC is warning that data center demand is raising the risk of winter blackouts across the country ^[7]. Meanwhile, local opposition has blocked nearly $100 billion in projects, Texas's governor has halted some 1,800 facility approvals, and New York has imposed a statewide moratorium ^[8][9]. None of these were designed as safety mechanisms. They are just what happens when a buildout runs into a wall. And the federal government is now working against the only brakes that exist. The House passed a budget bill that would freeze state and local AI regulations for ten years ^[10]. The president is touring communities to sell the buildout as a windfall.

Your community and your state is going to become rich and there's very little disruption. — Donald Trump

The industry itself cannot agree on whether it wants to slow down. OpenAI's chief scientist says the work should be slowed and that some international norm is needed — while his own company is racing to fully automate its AI researchers by 2028 ^[11].

We actually believe this should be slowed down … We need some sort of international norm to be able to control this. — Jakub Pachocki

Anthropic's co-founder says the opposite.

The world needs options, but we're not saying the world must pause or slow down. That's not what the evidence says. — Jack Clark

Yet both firms are still asking Washington for the same containment-era tools — third-party pre-release testing, a FINRA-style oversight body — the very mechanisms that just failed ^[12][13]. The result is a slowdown nobody designed and nobody controls. It is not a safety framework; it is a power line and a zoning board. No agency can claim credit for it, and no one is in charge of it. That is what governing AI looks like now.