OpenAI and Anthropic Are Writing Their Own AI Rules

On June 2, 2026, President Trump signed Executive Order 14409, directing federal agencies to design a framework under which AI developers could give the US government early access to certain models before releasing them publicly. The order set a 60-day clock for agencies to define what counts as a “covered frontier model” and how the review would work. That clock runs out on August 1 — this Saturday. Over the past two weeks, the White House has been circulating drafts of that framework to OpenAI, Anthropic, and Google, and all three have been sending edits back.

What the order actually does

EO 14409 doesn’t mandate anything on its face. It asks agencies to build a voluntary process: a classified benchmarking regime, run under the NSA Director’s authority, that screens AI models for advanced cyber capabilities and flags the ones judged to pose a national-security-relevant risk. Once a model is designated “covered,” the framework would give the federal government up to 30 days of pre-release access to test it before the developer ships it publicly. The order itself doesn’t define which models qualify — that definition is exactly what’s supposed to get nailed down by the August 1 deadline, through the classified benchmarking process rather than through public rulemaking.

That’s a lot of load-bearing detail left to a deadline most people haven’t heard of, for a process most of the public will never see the inside of.

Why OpenAI and Anthropic are in the room

The two companies aren’t just commenting on the framework from outside — they’re reportedly editing the actual draft text, and pushing specifically for the standard to apply across the whole industry rather than only to labs that already cooperate with Washington. On its face that sounds like the more responsible position: if a review threshold exists, better it catch everyone with a sufficiently capable model, including Meta and xAI, than let competitors skip it by staying uncooperative. OpenAI and Anthropic have framed it that way publicly — a consistent standard instead of a patchwork.

But it’s worth sitting with the structural oddity here regardless of intent: the companies with the most resources to absorb a 30-day compliance review, and the most institutional relationships with the agencies writing it, are helping decide what counts as a model serious enough to need one. A threshold built around what OpenAI and Anthropic’s safety teams already understand and measure is a threshold that’s comparatively easy for OpenAI and Anthropic to clear. Whether that’s regulatory capture or just the unavoidable reality that the government doesn’t have in-house expertise to write cyber-capability benchmarks without the labs’ help is a genuinely open question — probably some of both.

The voluntary word is doing a lot of work

The sharper criticism isn’t about who’s drafting the framework — it’s about what “voluntary” will mean once it exists. Critics point to a precedent already on the books: the administration has separately imposed export controls on Anthropic’s models, a clear instance of the government using regulatory leverage against a specific lab rather than asking nicely. Against that backdrop, a “voluntary” pre-release review that a company can technically decline looks a lot like a review a company can decline the way you can decline a base rate increase — allowed in theory, costly in practice. If declining the review becomes a signal that invites closer scrutiny elsewhere, or if government contracts and goodwill start flowing to labs that comply, then the distinction between voluntary and mandatory collapses without a single new law being passed.

Disputes over the current draft reportedly center on exactly this kind of scope question: what counts as a “frontier model” in the first place, whether open-weight models get carved out, and how much teeth “voluntary compliance” actually has. None of that is settled yet, which is part of why the August 1 deadline matters — it’s the point where vague executive-order language turns into an operational definition that either does or doesn’t apply to the next model any of these companies ship.

How this sits next to yesterday’s letter

We wrote yesterday about the “Pacing the Frontier” letter, where more than 1,100 employees at these same companies asked the government to build tools for a future, internationally coordinated slowdown mechanism — something that doesn’t exist yet and might never be needed. This week’s story is the unglamorous, already-happening counterpart: a current government review process, with a real deadline this week, that the companies are shaping from the inside rather than petitioning from the outside. It’s a useful contrast. Asking Washington to build hypothetical brakes for later is easy to endorse and costs a company nothing today. Negotiating the actual scope of a real review process that could delay your own model launch is where the incentives get complicated, and it’s the negotiation that’s happening quietly this week while the letter gets the headlines.

None of this means the framework is bad policy, or that OpenAI and Anthropic are acting in bad faith by participating — somebody has to define “covered frontier model,” and the labs plausibly understand model capabilities better than anyone else in the room. But the right way to evaluate a self-drafted oversight regime isn’t to ask whether it sounds reasonable. It’s to watch what the August 1 definition actually excludes, and whether “voluntary” ever gets tested by a lab that says no.

Sources: TechTimes, crypto.news, Digitimes, Congress.gov CRS report on EO 14409, TechTimes on the August 1 deadline