Meta is notably absent from the White House's landmark AI safety agreement, even as the government prepares to formalize its first-ever framework for reviewing frontier AI models before release. While OpenAI, Anthropic, Google, Microsoft, and Amazon negotiate the terms of a voluntary pre-release review process, Meta has stayed outside the deal, raising questions about how the government plans to oversee open-weight models like Llama that are freely distributed to developers worldwide.

Why Is Meta Missing From the White House AI Deal?

The White House's August 1 deadline marks the expiration of a 60-day review period set by President Trump's June 2 executive order on AI safety. The framework being negotiated involves the five largest AI companies, but Meta's exclusion is conspicuous. The government's approach to frontier models has centered on controlling access before release, a strategy that works well for proprietary systems like OpenAI's GPT-5.6 Sol or Anthropic's Claude Fable 5. But Meta's Llama models operate under a fundamentally different distribution model: they are open-weight systems released publicly for anyone to download, modify, and deploy.

This distinction matters enormously for government oversight. When Anthropic released Claude Fable 5 and Mythos 5 in mid-June, the Commerce Department suspended global access within roughly a day, citing national security concerns related to a jailbreak that exposed the models' advanced cyber-offense capabilities. The company complied, and access was restored three weeks later after Anthropic agreed to work with Amazon, Microsoft, and Google on shared security standards. That enforcement mechanism relies on centralized control over distribution, something that doesn't exist for open-weight models once they're released into the wild.

What Does the Government Actually Control Under the New Framework?

The executive order's language promises that nothing in it creates "a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models." But that description applies only to the voluntary framework itself. The government's actual toolkit is far broader. The Export Control Reform Act of 2018 gives the Commerce Department authority to establish controls on emerging technologies deemed essential to national security, without needing a new statute or executive order.

The practical implication is stark: the word "voluntary" describes what the executive order does, not what the government can do. Before the formal framework even existed, the Commerce Department used export control authority to suspend Anthropic's models globally. The White House separately asked OpenAI to limit GPT-5.6 Sol's rollout to government-vetted partners. Both companies complied without legal challenge. That is the operating framework that August 1 will formalize.

For enterprise AI buyers and developers, the question is not whether to participate in the voluntary framework. The question is whether the government will treat non-participation as a commercial or national security matter, because it already has the tools to respond in either register.

How the Government Plans to Identify "Covered" AI Models

The executive order gives three agencies, the NSA, CISA, and the Treasury Department, 60 days to produce two critical deliverables: a classified benchmarking process that determines which AI models qualify as "covered frontier models," and a voluntary framework governing how developers engage with the federal government before release. The NSA Director holds sole authority to designate which models qualify, and the criteria are classified and will remain classified.

What developers and the public will know after August 1 is the process for submitting models and the general structure of the review. They will not know exactly what capability level triggers the designation. A developer could inadvertently cross the threshold and trigger a 30-day review window without warning. The benchmarking process is understood to assess a model's autonomous cyber capabilities, specifically its ability to find and exploit weaknesses in software on its own.

The UK's AI Safety Institute, working with the Five Eyes intelligence alliance, has been tracking frontier model cyber capability and documented a model succeeding 73% of the time on expert-level capture-the-flag cybersecurity challenges, a tier no model could clear before April 2025. In a joint statement issued June 23, the Five Eyes grouping warned that AI models capable of launching major cyberattacks that could overwhelm government and enterprise defenses are "months, not years, away".

Steps to Prepare for the New AI Safety Framework

• Understand Export Control Authority: Organizations developing or deploying frontier AI models should recognize that the Commerce Department can apply export controls to AI models themselves, not merely to their weights or source code, under authority that predates the executive order by eight years.
• Monitor Jailbreak Severity Standards: The five labs participating in TRAINS, the Testing Risks of AI for National Security program, are developing a shared jailbreak severity scoring system modeled on the Common Vulnerability Scoring System used in software security, which will establish a common language for assessing model vulnerabilities.
• Prepare for Potential Review Delays: Companies planning to release frontier models should anticipate that a 30-day review window could be triggered without advance notice of the specific capability thresholds that trigger designation, requiring flexibility in product launch timelines.
• Engage With Voluntary Standards: Even though the framework is nominally voluntary, the precedent set by the Anthropic and OpenAI cases suggests that non-participation may be treated as a national security matter, making proactive engagement with government reviewers a practical necessity.

Brad Carson, head of Public First, a bipartisan pro-AI safety organization, captured the tension accurately after the Anthropic episode. "The Fable episode shows the need for clear regulations. Right now, you have an ad hoc, personalized, opaque, possibly lawless approach," he told CNN.

"The Fable episode shows the need for clear regulations. Right now, you have an ad hoc, personalized, opaque, possibly lawless approach," Carson stated.

Brad Carson, Head of Public First

Meta's absence from the negotiations raises a critical question: how will the government apply this framework to open-weight models that are designed to be distributed freely? The answer will likely shape how the AI industry approaches model releases for years to come. If the government treats open-weight models as outside the voluntary framework, it may signal that the export control authority will be the primary enforcement mechanism. If Meta's Llama models eventually trigger that authority, it could establish a precedent that fundamentally changes how open-source AI development operates under government oversight.

The August 1 deadline will reveal the formal structure of the voluntary framework, but the more consequential question, the one already answered by the events of June and July, is what "voluntary" actually means when the government has unilateral authority to suspend access to any AI model it deems a national security risk.