Белый дом исключил открытые AI-модели из нового фреймворка кибербезопасности

The White House published on August 4, 2026, a voluntary framework for assessing cyber risks from advanced AI systems — and immediately excluded all open models from it, Axios reports. The tool, created in fulfillment of President Donald Trump's executive order, explicitly prohibits its application against open AI systems whose weights are available for free download.

Where the framework came from

President Trump signed the executive order in June 2026, obligating AI companies to provide the federal government with access to their frontier models before public release — for cybersecurity threat assessment. About two months later, the White House presented the framework itself.

Key parameters of the document according to Axios:

• Published in early August 2026 — two months after the presidential executive order
• Voluntary in nature: participation for AI companies is not mandatory
• Focused exclusively on closed commercial AI systems
• Explicitly states that it cannot be applied to restrict already-released open models

Why open models were left outside the scope

Open models — such as Meta Llama or Mistral AI developments — are by their nature unavailable for centralized control. Their weights are published in the open: any user can download the model, run it locally, and work with it without the developer's knowledge. Once a model is released — it is impossible to put it "back under the lid."

The framework's authors, according to Axios, did not simply leave open models outside the document's scope — they explicitly stipulated: the tool cannot be used to restrict or regulate already-published open systems. This formally closes any loophole for future restrictions through this mechanism.

Notably, even for closed laboratories formally falling under the framework's scope, participation remains voluntary — the document provides for no mandatory testing.

According to the description cited by

Axios, the White House itself calls the document "voluntary guiding principles" — without any enforcement mechanisms.

What the framework changes in practice

The American AI risk assessment system covers only part of the market — closed commercial laboratories: OpenAI, Google DeepMind, Anthropic. All of them already maintain dialogue with federal structures, so the framework consolidates an already established status quo.

The open AI community remains completely outside any federal oversight framework. It is precisely open models that often find themselves at the center of safety discussions: they cannot be recalled after publication, forcibly updated, or blocked on a geographical basis. The framework does not resolve this issue — and, judging by the text, did not intend to.

The Verge characterized the document as "limited and vague" — which accurately reflects its legal status: voluntary, narrow in scope, and devoid of enforcement mechanisms.

What this means

The first American government framework in the field of AI safety records not regulation, but its absence: dialogue with closed laboratories on a voluntary basis, the open AI space — untouched. This choice is now enshrined in an official document.

*Meta is recognized as an extremist organization and is banned in Russia.