AI Policy

Five AI Labs Know What Washington Won't Say

Executive Order 14409 gave the US government 60 days to design a frontier AI review framework. The August 1 deadline passed with no public text, while OpenAI, Anthropic and Google have already seen the draft. Here is what that gap means for any EU business choosing an AI vendor.

AI PolicyBy Servola Tech Desk2026-08-054 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Five AI Labs Know What Washington Won't Say

Key takeaways

Executive Order 14409 gave the US government 60 days to design a voluntary review framework for frontier AI models. That design deadline, 1 August 2026, passed with no Federal Register notice, no NIST or CISA publication, and no statement from the White House's own science office.
Drafts of the framework had already been shown to OpenAI, Anthropic and Google by 27 July, and described as close to final, while every competitor, every customer and the public have seen nothing.
The framework would ask labs to hand the government up to 30 days of advance access to a covered frontier model before its public release, but nobody outside those three companies yet knows which models that word covers. The order leaves the definition to the NSA Director, case by case.
The order contains no provision at all for open-weight models. Anthropic has publicly argued that any sufficiently capable model should face the same review regardless of how it is distributed, and that question is still unresolved.
If you are choosing between a closed API from OpenAI, Anthropic, Google, Microsoft or xAI and a self-hosted open-weight model from Mistral, DeepSeek or Meta, only the first group sits anywhere near a government review queue whose rules are still being written without you, or your market, in the room.

A Deadline That Passed Without A Sound

Executive Order 14409 asked three parts of the US government to build a frontier AI oversight system in 60 days: a classified benchmarking process run jointly by the NSA, CISA and NIST, a voluntary disclosure framework coordinated by Treasury, NSA, CISA and NIST, and a federal cyber workforce expansion plan from the Office of Personnel Management. All three had the same design deadline, 1 August 2026. It came and went. No Federal Register notice appeared. NIST and CISA published nothing. The Office of Science and Technology Policy, the White House unit meant to coordinate the effort, issued no statement at all.

None of this makes the framework illegal or void; the order set an internal design deadline for the government, not a compliance date for companies. But a 60-day mandate that produces no public output on day 61 tells you something about how contested, or how unfinished, the underlying policy still is.

Five Labs Already Know The Draft

The silence in public does not mean nothing happened in private. Reporting from 27 July, days before the deadline, described draft framework text already circulating to OpenAI, Anthropic and Google, with the mechanics close to final: labs would give the government up to 30 days of advance access to a covered frontier model before its public release, in exchange for early input on how the review runs. The order is explicit that this creates no mandatory licensing, preclearance or permitting requirement. It is, on paper, voluntary.

What is not voluntary is the information gap it creates. Three companies have read language that the rest of the market, including every European customer deciding which of their products to build on, has not. Separately, Axios has reported that OpenAI, Anthropic, Google, Microsoft and xAI are the named participants in a related effort, TRAINS, to standardize how severely a jailbreak or exploit should be scored, another process running ahead of any public framework text.

Why It Matters That Nobody Knows The Definition Yet

Why it matters: the single most consequential word in the whole framework, covered, has no public definition. The order leaves that determination to the NSA Director, informed by candidate measures such as model size, deployment mode and third-party capability evaluations. Until that threshold is set, no outside business, and arguably no lab outside the three that have seen the draft, can say with confidence whether a given future model release will trigger a 30-day government review window or not.

Yes, but: the order is careful to disclaim a licensing regime, and Anthropic itself has publicly pushed for the framework to go further, applying to sufficiently capable models regardless of whether they are closed or open-weight. That is a genuine safety argument, not empty positioning. It does not change the fact that, for now, open-weight releases sit outside any version of this scheme that has been described.

The bottom line: a policy that is voluntary in name, drafted behind closed doors with three companies, and silent on its own core definition, is not yet a rule you can plan around. It is a live negotiation between Washington and its biggest AI vendors, and the outcome will land on whichever companies end up inside the covered-model definition, whenever that definition finally appears.

The Instruction For Anyone Betting On A US AI Vendor

If your business runs on an API from OpenAI, Anthropic, Google, Microsoft or xAI, treat this framework as a live variable in that vendor relationship, not settled background. Ask your vendor directly whether a future model your product depends on could fall under a 30-day pre-release review once the covered-model definition is finalized, and what that would mean for your release timelines if it does. If instead you run a self-hosted open-weight model from Mistral, DeepSeek or an open Llama release, you are outside this framework as written today, but that gap is exactly what Anthropic is lobbying to close, so do not treat it as a permanent feature of your stack. Either way, the deadline that just passed in silence is not the end of this story. It is the government telling you, without saying a word, that the terms are still being decided without you.

Frequently asked questions

What is Executive Order 14409?

It is the US executive order that commissioned a voluntary review framework for frontier AI models, along with a classified cyber-capability benchmarking process and a federal cyber workforce expansion plan. It set a 60-day internal deadline, 1 August 2026, for the government to design these programs.

Did the August 1 deadline mean AI companies now have to comply with something?

No. The deadline was for the US government to finish designing the framework, not for companies to comply with it. No Federal Register notice, NIST or CISA publication, or White House statement appeared by that date, and the order explicitly disclaims creating a mandatory licensing or preclearance requirement.

Are open-weight models like Mistral's or DeepSeek's covered by this framework?

As written, the order contains no provision addressing open-weight models at all. Anthropic has publicly argued they should face the same review if sufficiently capable, but the government has not resolved that question, so open-weight releases currently sit outside any version of the scheme that has been described.

A rule drafted with three companies and silent on its own definition is not yet a rule. It is a negotiation, and the deadline that passed without a word is the clearest sign yet of how much is still being decided behind that silence.

Executive Order 14409Frontier AI regulationOpenAIAnthropicOpen-weight models

Google Scrapped a Near-Finished AI Model

Google DeepMind scrapped a nearly finished Gemini base model and moved the launch to 17 July. The sunk-cost discipline behind the call, and the trap it sets for buyers.

3 min read

Cybersecurity

Is Your Cyber Defense Ready for Machine-Speed Attacks?

OpenAI launched Daybreak and Anthropic has Project Glasswing: both leading labs now ship autonomous AI cyber defense. The real signal is that attacks already move at machine speed, and the winners will be the ones who deploy AI security with real oversight.

2 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.