Meta Is Washington's Last Holdout on Frontier AI

The Trump administration is privately pressing Meta to submit its frontier AI models for federal security review — making Meta the only major U.S. developer that has not agreed to share unreleased models with Washington, per The New York Times and Reuters. OpenAI, Anthropic, Google DeepMind, Microsoft, and xAI already signed on. Ten days after Trump called the program voluntary, Commerce ordered Anthropic to kill Fable 5 and Mythos 5 worldwide. Voluntary is the official word. Pressure is the practice.

What's new: On June 23, four people familiar with confidential administration emails told the NYT that Washington wants Meta inside the Center for AI Standards and Innovation (CAISI) review process — the federal program that tests frontier models for cyber, biosecurity, and agentic risks before public release.

• Meta launched Muse Spark — its first closed-source frontier model — on April 8, breaking from the open-weight Llama lineage.
• The model now powers Meta AI across the company's apps; larger Muse models are in development.
• Meta spokesman Francis Brennan told Reuters: "We share the administration's goal of advancing U.S. leadership on robust and secure frontier AI. While we are working through the details, we hope to sign the agreement soon."
• Commerce spokesman Ben Kass told the NYT: "This story is not unusual. It is the very work CAISI is supposed to be doing."

Why it matters: CAISI is becoming the gate between frontier labs and sensitive buyers — defense, finance, healthcare, critical infrastructure. A model without a federal review stamp may soon look like a model without a security clearance.

The voluntary framework: On June 2, President Trump signed Executive Order 14409, directing agencies to design a voluntary program giving the federal government up to 30 days of access to "covered frontier models" before labs release them to other trusted partners.

• Section 3(c) explicitly states the order does not authorize mandatory licensing, preclearance, or permitting for AI model release.
• NSA, CISA, and NIST are building classified benchmarks to decide which models qualify as "covered."
• Google DeepMind, Microsoft, and xAI joined OpenAI and Anthropic in May agreements to provide pre-release access, per Reuters.

The other playbook: Ten days after the voluntary order, Commerce issued a mandatory directive requiring Anthropic to block foreign nationals from Fable 5 and Mythos 5 — anywhere in the world. Anthropic disabled both models globally within hours.

• Legion LegalTech, a San Jose litigation-tech startup, sued the government on June 23 over the collateral damage to its Canada-based developers.
• The contrast is stark: five labs negotiated voluntary access. One lab got an export-control kill switch. Meta is negotiating while watching both outcomes.

Why Meta might resist: Meta spent years building developer goodwill through open-weight Llama. Muse Spark reversed that bet — closed weights, proprietary architecture, Alexandr Wang's Superintelligence Labs rebuilding the stack from scratch.

• Handing unreleased weights to a classified federal review team means exposing the competitive secrets Meta just spent nine months locking down.
• Meta's own security week makes the optics brutal: on June 22, the company paused mandatory employee keystroke tracking after sensitive internal data became visible companywide.
• Washington is asking a company that cannot secure employee screenshots to open its frontier models to NSA benchmarking.

What Meta gains by waiting: Time. Every week outside CAISI is another week Muse Spark ships to billions of users without a classified red-team report attached.

• Meta's statement — "working through the details" — is diplomatic code for negotiating IP protections, access scope, and which models count as "frontier."
• Open-source advocates have long warned pre-release review regimes could entrench incumbents — a tension Meta navigated for years as Llama's champion.

What Meta risks: Procurement gravity. Federal buyers, defense contractors, and regulated enterprises increasingly treat CAISI participation as a vendor qualification signal.

• A two-tier market is forming: reviewed models for sensitive deployment, unreviewed models for everyone else.
• Meta's social-platform distribution is enormous — but enterprise and government contracts are where AI revenue gets priced for IPO season.
• Continued holdout invites the same political pressure Anthropic faced — except Meta has not yet filed suit against anyone.

Convina's view: Washington sold the industry a voluntary framework and demonstrated a mandatory one in the same month. Meta is the last major lab still pretending the choice is real. Signing CAISI access means trusting the government with the proprietary weights Meta just bet its future on. Refusing means betting Washington will keep sending emails instead of letters from Commerce. Neither option is clean — and Meta's internal security failures this week make the trust deficit mutual. The frontier AI market is not waiting for a mature rulebook. It is sorting vendors into two columns: those who showed Washington their models before launch, and those who learned what happens when they don't. Meta has days, not quarters, to pick a column.