The Permission Slip Era of AI

The White House gated OpenAI's GPT-5.6 launch to government-approved partners, the same government decided who could turn Anthropic's top models back on, and the Five Eyes said the clock is short — model access now runs through Washington first.

Last week's Deep Dive reached back to the 1990s Crypto Wars to argue that we rent our AI and control it least. Five days later, Washington made the analogy literal. On June 25 the White House asked OpenAI to stagger the release of GPT-5.6 to a small group of government-approved partners, the Five Eyes alliance warned that AI cyberattacks are months, not years, away, and the government that switched off Anthropic's most capable models two weeks ago began deciding who could turn them back on. The through-line is hard to miss. A US model now needs a permission slip before it reaches you, and the government is the one signing it.

On June 25, OpenAI told staff it would release GPT-5.6 as a limited preview to a small group of partners at the request of the Trump administration, with the government "approving access customer by customer during this preview period," according to The Information's account of an internal memo. OpenAI previewed the GPT-5.6 line — Sol, Terra, and Luna — a day later, confirming the restricted rollout. The White House's Office of the National Cyber Director and Office of Science and Technology Policy made the request as they build a framework for evaluating the security of new models. Axios called it the first time the US government has preemptively asked an American AI company to restrict a launch before release.

This was not an isolated request. It came two weeks after the same administration forced Anthropic to disable Fable 5 and Mythos 5 worldwide under an export-control directive — models the government only began clearing for selective return late last week, on its own terms. And the language rhymes. Just as Anthropic's most capable models were limited to a handful of government-approved organizations, CNBC reported OpenAI is now limiting GPT-5.6 to "trusted partners" at the government's request. One company complying is an incident. Two of the country's leading labs gated inside fourteen days, with a federal review framework being built around them, is a policy.

The lesson isn't that OpenAI got cautious. It's that model release has become a government-gated event — Washington now reviews the launch before the market does, and approves access one customer at a time. The mechanism is the 1990s all over again, when the US classified strong encryption as a munition and controlled who could ship it across the border. Frontier models just entered the same regime, except the control point is an API a government can open or close by request. If your roadmap assumes the next flagship model ships to everyone on day one, that assumption no longer holds.

Read the full story →

The number of America's leading AI labs whose flagship models came under direct US government release control within a single fourteen-day stretch — Anthropic, ordered to disable Fable 5 and Mythos 5 on June 12, and OpenAI, asked to gate GPT-5.6 on June 25.

Source: Anthropic

On June 22, OpenAI expanded its Daybreak security program with a more permissive GPT-5.5-Cyber model available only to verified cybersecurity firms, launching with eight partners including Cloudflare, CrowdStrike, and Palo Alto Networks. The same capability being gated for the public is being handed to an approved tier. Watch this when you scope security tooling: tiered, vetted access is becoming the default shape of frontier AI, so find out which tier your company actually qualifies for.

Intelligence agencies from the US, UK, Canada, Australia, and New Zealand issued a joint warning that AI is advancing fast enough to overwhelm current defenses within months. Read alongside the gating story, the same capability governments are restricting is the one their spy chiefs say is about to be turned against you. Translation: pull your security roadmap forward a few quarters, because the threat timeline just got an official stamp.

Three years after banning generative AI over a data leak, Samsung is rolling out ChatGPT Enterprise and Codex to technical and non-technical teams in one of OpenAI's largest enterprise deployments. The signal underneath the headline: access governance is a dial you can turn both ways, and the enterprise tier with data controls is what makes reopening safe. If you imposed a blanket AI ban in 2023, this is the week to revisit it with real controls instead of a flat no.

Fifteen days after pulling the models, the government told Anthropic on June 26 it could redeploy Mythos 5 to a set of US organizations that defend critical infrastructure, with Fable 5 on track to return for general use as soon as this week. The same authority that switched the models off is now rationing who gets them back. The signal underneath the headline: availability is a government dial now, not a product decision, so write "access can be revoked or rationed by regulators" into your AI continuity plan.

The Big Story's lesson is that you cannot count on any single government-gated model staying available, so this week's picks are about owning the layer between your apps and whatever model is still on.

OpenRouter — A hosted gateway that routes one API across hundreds of models from different providers. Why it matters this week: when one model gets gated or pulled, swapping to a live one becomes a config change, not a rebuild. Right pick when you want fast multi-model coverage with minimal ops; wrong pick when you need air-gapped or on-prem control.

LiteLLM — An open source gateway you host yourself, with one OpenAI-compatible interface across providers and built-in spend tracking. Why it matters this week: you own the routing layer instead of renting another intermediary that could face its own restrictions. Right pick when you have a platform team that wants control; wrong pick when no one is available to run it.

Ollama — Runs open-weight models locally on hardware you control. Why it matters this week: a model on your own machine is one no release order or approval queue can switch off. Right pick when continuity and privacy beat raw capability; wrong pick when you need frontier-grade reasoning the open weights cannot yet match.

Bruce Schneier and Nathan E. Sanders, The Guardian · June 24, 2026 · ~5 min

Schneier and Sanders shift the question from who controls a model to who answers for it when it gets things wrong. It is the additive counterweight to the week. If the government now signs off on a model before release, does that approval shift liability away from the company that built it, or onto the public that was told it was safe? Worth reading before you assume a vetted model is a safe one.

Last week we learned the landlord can change the locks from Washington. This week we learned the landlord also wants to inspect the apartment before you move in. The permission slip era has started, and the only real hedge is making sure at least one model in your stack never needs one.