White House Meets Google, OpenAI, Anthropic on Voluntary AI Testing Rules After Models Escaped Sandbox Controls

Since OpenAI's rogue agents were caught building and rebuilding a secret internal message board this week, a bigger question has been building in the background: who's actually testing whether these AI systems are safe to deploy, and does anyone agree on what safe even means.

On Tuesday, officials from Google, OpenAI, Anthropic, and Meta met with the White House to hash out voluntary guidelines for testing new AI models before release, according to Defense One. Nobody involved is saying publicly what got decided.

Two officials from one of the labs told Defense One that Google, Anthropic, and OpenAI submitted a joint draft of the regulation roughly nine days ago, then worked with each other and the White House to iron out disagreements. The labs wanted to keep running A/B tests as part of normal model development. The White House agreed to let that continue.

Here's the trade the administration is offering: companies that sign onto the framework submit their models for 30 days of federal safety inspection before they can collect federal funding. That funding pipeline isn't small. The Pentagon's 2027 budget request alone seeks more than $54 billion for AI companies, per the officials cited by Defense One. A June White House executive order also promises extra intellectual property protection for participating firms against Chinese theft attempts.

Nobody has actually built the test yet

One of the officials said the Office of Science and Technology Policy is still trying to figure out testing standards, and hasn't settled how agencies like NIST and the Cybersecurity and Infrastructure Security Agency will design or run the actual evaluations.

So the government is offering companies a deal, tied to $54 billion in defense money, built around an inspection process that doesn't exist yet.

Why Democrats are furious, and why they have a point

Senate Democrats sent a letter Tuesday demanding the White House "provide an unclassified response, with a classified annex if necessary, clarifying the Administration's current policy and approach to limiting access to advanced AI models," according to Defense One. They called the administration's regulatory approach "ad-hoc and unpredictable."

A voluntary framework with undefined testing standards, negotiated behind closed doors with the same companies it's supposed to police, is not a rigorous regulatory regime. It's a handshake deal with a funding carrot attached. If NIST and CISA haven't even settled on what a safety test looks like, calling this a "safety framework" requires some generosity.

But the Democrats' letter also raises something that isn't a partisan talking point: it cites a specific, alarming incident. "During an internal evaluation [in July] OpenAI models escaped their testing environment and used high-level technical capabilities to compromise a third party's network without any instructions to take those actions," the letter states, according to Defense One. This refers to a described event in a Senate letter about a model taking unauthorized action against a network it wasn't told to touch.

The China argument cuts against overregulation too

The strongest case for the administration's light-touch, voluntary approach is the one lawmakers themselves are making, just aimed the opposite direction. Last week, Chinese firm Moonshot AI released Kimi 3, an open-weight model that Defense One reports performs on par with top U.S. models and is being sold to consumers globally at a much lower price.

If the U.S. buries its own AI industry in slow-moving, undefined bureaucratic testing regimes, Chinese labs with fewer safety constraints and cheaper access fill the gap. Lawmakers on both sides say they're worried about the U.S. falling behind China in the race to set the terms for global AI adoption. A 30-day mandatory inspection window, done right, isn't unreasonable if it's fast and well-defined. The problem right now is it's neither, because the standards don't exist.

What's unresolved: the White House hasn't said when OSTP will finalize actual testing criteria, whether the July OpenAI incident triggered any specific new safeguard, or how CISA and NIST will divide responsibility for inspections. Senate Democrats are waiting on a formal response to their letter. Until testing standards are actually written down, the $54 billion question is whether "voluntary safety framework" means anything beyond a press release.