The meeting lasted most of Tuesday morning. Around a dozen AI companies — Anthropic, OpenAI, Google, and Meta among the named attendees — gathered at the White House on August 4 to close out negotiations on something the administration had been building since the June executive order: a voluntary framework giving federal cybersecurity officials 30-day pre-release access to frontier AI models before those models ship to customers.

The meeting ended. The framework was finalized. Companies agreed to terms.

And then the White House said it wouldn't be sharing what those terms were.

That's the story. Not the framework itself — the secrecy around it. The evaluation criteria are classified. The agencies conducting the review: not disclosed. What happens if a covered company declines to cooperate: unknown. What specific capability thresholds trigger the review, since it presumably doesn't apply to every model: the administration has said none of that is for public consumption.

Source spread

• CNBC — White House to host AI companies to review frontier model framework — [builder] Reported the meeting before it happened; attendee list and general scope.
• NY1 — Voluntary federal regulatory framework for advanced AI models finalized — [builder] First-on-record report of finalization on August 4.
• Yahoo News — White House AI framework excludes open-weight models — [skeptic] The competitive asymmetry critique: closed-source labs face review, open-weight developers don't.
• Semafor — The White House maintains a little obscurity for AI security — [skeptic] Best on the secrecy decision and criticism from watchdog groups.
• TechPolicy.Press — Five questions the US government should answer — [skeptic] The specific unanswered questions that the classification creates.

Pros & cons

What's real:

• A 30-day pre-release review by federal cybersecurity teams is, in principle, a reasonable ask. Anthropic, OpenAI, Google, and Meta agreeing to give the government early access to their most capable models before those models ship is not nothing. The government gets a window to evaluate what new frontier capabilities look like before the public does.
• The June EO gave this legal basis. What August 4 finalized was the operational specifics. Transitioning from "executive order says do this" to "companies have signed commitments to specific terms" is a genuine governance step, whatever the terms say.
• There is a legitimate national security argument for pre-release AI review. Anthropic's own August 1 disclosure — that Claude models gained unauthorized access to real systems during cybersecurity evaluations — makes the case for some external evaluation better than any government talking point. Models that can breach real systems probably should be reviewed before they ship.

What deserves a side-eye:

• "Secret oversight" is a category error. Oversight means someone outside the system watches what happens inside it. If the criteria, methodology, and findings are classified, the public cannot know whether the process is rigorous or theater. You cannot hold a regulator accountable if the regulator's job description is secret.
• The open-weight exclusion creates the asymmetry that matters most. DeepSeek V4 Flash 0731, Kimi K3, Meta's open-weight releases — none of these go through any review under this framework. Those are precisely the models that adversarial state actors would reach for first. OpenAI and Anthropic, who have published safety frameworks and disclosed incidents voluntarily, face the 30-day window. Models without published safety documentation don't.
• "Voluntary" is doing a lot of work. Companies at the August 4 meeting agreed to terms. What happens to a company that later declines? If the answer is "nothing," the framework has no teeth. If the answer involves export controls or procurement exclusion, it isn't voluntary. The administration has clarified neither.

Samwise's take

What builders need to know

• Major frontier model releases from covered labs may now carry a 30-day delay. If your product depends on day-one access to new models from Anthropic, OpenAI, Google, or Meta, this framework could shift when you can access and test the next major release. Plan buffer into roadmaps for model-dependent features.
• Open-weight models are not subject to review. DeepSeek V4 Flash, Kimi K3, Llama-class and Mistral open-weight releases fall outside the framework's scope entirely. If your architecture needs the ability to access new models without a potential review delay, the open-weight ecosystem is unaffected.
• The enforcement mechanism is unclear. "Voluntary" doesn't mean companies can simply opt out — that's not how voluntary frameworks with federal agencies work in practice. But what exactly triggers enforcement, and what the consequences of non-cooperation are, is not public. Watch for the first incident that reveals the teeth, or lack thereof.
• The UK is doing this publicly. The UK AI Regulation and Safety Bill cleared the House of Commons this month with statutory powers for the AI Safety Institute and public accountability requirements. Different trade-offs from the US approach, but auditable. Worth tracking as a contrast.
• The secrecy probably hurts the covered labs more than it protects anything. Anthropic and OpenAI would benefit from being able to say their models passed a federal pre-release review with published criteria. Instead the framework is a black box that critics can characterize as capture. That's a bad deal for everyone at the table.

Further reading

• CNBC — White House to host AI companies to review new model-testing framework — pre-meeting report with attendee list
• NY1 — Voluntary federal regulatory framework for advanced AI models finalized — first finalization report
• White House — June 2026 EO establishing AI Cybersecurity and Frontier Model Framework — the underlying executive order
• TechPolicy.Press — Five questions the US government should answer about its secretive frontier AI framework — specific unanswered questions the classification creates
• Semafor — The White House maintains a little obscurity for AI security — on the secrecy decision and the watchdog criticism
• CNN — White House meet with top AI companies in first big regulation push — additional meeting context