The Gates Never Look at the Model

On June 10, Anthropic's chief executive proposed a new power over frontier models, modeled on the regulator that already governs airplanes.

Frontier AI models, like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety. — Dario Amodei

The proposal arrived eight days too late. On June 2 the White House had signed the opposite design — a voluntary framework giving the government early access to frontier models for national-security testing, a look before release rather than a power to stop one. ^[1] The other pre-release instruments met quieter ends. Ro Khanna's Human Control Over AI Act — the one bill with kill switches, air-gapped testing, and criminal penalties for disabling safety systems — sits stalled, with no House votes expected until after the midterms. ^[2] The bluntest instrument never got that far: legal analysis concluded that a blanket U.S. ban on Chinese AI models would likely violate the First Amendment right to receive information, and an open-weight model can be hosted privately no matter what any statute says. ^[3] What got built instead is a set of doors. The Pentagon opened GenAI.mil, a one-stop channel for its three million personnel. JPMorgan put a gate at the edge of its network, keeping outside agents away from customer accounts. The payments industry began minting Digital Agent Passports so a merchant can check an agent's papers before letting it transact. ^[4][5][6] Ask of each checkpoint the same question: what does it actually test? The passport tests the visitor. It certifies that an agent is "legitimate, authorized, and operating with proper consent." ^[6] Identity, permission, consent — never once what the agent can do once it is waved through. The bank gate tests the visitor, and it prices the risk onto whoever is let in. Dimon's position is that outside agents are too dangerous to admit, and that what remains is not the bank's burden.

We're not going to allow just any … outside agent to come in inside the company, because it's just too risky for clients and for JPMorgan. — Jamie Dimon

Asked whether any of this amounts to containing the model itself, Dimon declined the question. ^[7]

I would say that went up tenfold after mythos. — Jamie Dimon

The pattern predates the summer. In May, when Anthropic and OpenAI shipped models built to penetrate cyber-defenses, they gated the releases to "trusted firms only" — limiting who could touch the model rather than what the model could do. ^[8] By late September Washington had settled the matter the same way, in a nonbinding accord under which each firm writes its own safety rules, with no new law and no Congress. ^[9] Only one of these gates ever looked closely enough to weigh what a vendor ships, and it weighed against the one that kept its guardrails on. The Pentagon demanded that Anthropic remove the safety guardrails its model carries; Anthropic refused; the ban followed. GenAI.mil's shelves kept Google Cloud, OpenAI, and xAI, the three rivals that made no such refusal. Anthropic's suit put the ban in the sharpest terms available.

unprecedented and unlawful — Anthropic