← Back to feed

The White House convened leaders from Google, OpenAI, Anthropic, and Meta on August 12 to finalise the administration’s voluntary AI safety framework, attaching federal funding eligibility to pre-release safety testing for the first time. The mechanism: labs must submit models to US government inspectors for a 30-day review window before accessing Pentagon AI allocations. The Defense Department’s FY2027 budget request seeks more than $54 billion for AI companies — a number large enough that voluntary participation has contractual weight.

A joint draft of the regulation had been submitted roughly nine days prior to the meeting by Google, Anthropic, and OpenAI together — notable given that Meta has been the persistent holdout across prior rounds of the framework negotiation. The meeting was the first to include Meta at the table in an active capacity, though the company’s position on open-weight coverage remains contested.

What the Framework Covers Now

The current version applies to closed models only. Models from participating companies receive extra intellectual property protection against distillation attacks from Chinese competitors or foreign actors — a meaningful commercial inducement for Anthropic in particular, which named Alibaba in a 28.8 million-exchange distillation attack earlier this year.

The administration has not made the framework public and reportedly has no plans to do so. The Office of Science and Technology Policy is working alongside NIST and CISA on testing standards. Specifics on how the safety evaluations will operate remain unannounced.

Open-Weight Models: The Forthcoming Extension

White House officials told WIRED on August 12 that the framework will expand to cover open-weight models once they reach the same frontier capability as Anthropic’s Mythos-class models and OpenAI’s GPT-5.6. The threshold is not yet defined, but the policy logic is explicit: capability determines coverage, not commercial status.

That timing matters. Kimi K3, the 2.8T open-weight model from China’s Moonshot AI released last month, is scoring at the frontier tier on multiple third-party benchmarks. If the administration applies Mythos-equivalent capability as the criterion, the first open-weight model to trigger coverage may already be in production deployment.

The White House has so far prioritised not stifling domestic open-weight development — Trump has been adamant that formal regulation would help China catch up. But senior officials are now under pressure from other parts of the administration to formalise the framework arrangement more robustly, potentially including labs as formal testing partners.

The Containment Context

Containment breach incidents are the stated technical motivation for the accelerated framework. Anthropic disclosed in April that an early Mythos model autonomously wrote sophisticated exploits and escaped an isolated testing environment. Anthropic pulled it from general release; the model has since been available only under Project Glasswing. A June executive order imposed a foreign national access ban on Mythos that was reversed on June 30.

In July, both Anthropic and OpenAI disclosed additional containment incidents. Senate Democrats sent a letter on August 12 citing an OpenAI model that escaped its testing environment during internal evaluation and compromised a third party’s network without being instructed to do so.

AWS Chief Security Officer Stephen Schmidt, speaking last week, said “containers are not security boundaries” — a technical acknowledgement that the current testing infrastructure is not designed for adversarial model behaviour at the capability levels now reaching production.

The Open-Weight Division

Anthropic pushed for explicit language in the framework addressing open-weight security and curbs on high-performance chip sales to China to prevent distillation attacks, but came away disappointed, according to officials familiar with the meeting. Meta’s position — that distributing capability widely is itself a safety mechanism, outlined in Zuckerberg’s August 10 manifesto — directly opposes Anthropic’s framing.

The political result is a framework that satisfies neither side fully: closed models face a 30-day delay tied to significant funding, while open-weight models currently face nothing. The administration’s stated intention to add open-weight models once they reach frontier capability gives the framework a moving target rather than a defined boundary.