Administration officials told leading AI companies this week that open-weight models will be excluded from the government's new pre-release security review. The framework was set up by an executive order on 2 June and reviewed with Meta, Anthropic, Google and OpenAI at a meeting on 4 August. The exclusion was reported by Reuters, citing two sources.
What the framework is
Voluntary, and built around access rather than approval: the government gets up to 30 days with a company's most advanced model before release, to evaluate it for security risk. The evaluation methodology has not been published.
The logic of the carve-out
A model whose weights anyone can download and modify cannot meaningfully be gated by a pre-release window — the review would examine an artefact that is already unconditionally public. The counter-argument is that this is exactly why open weights need scrutiny, and it is the one Dario Amodei has made repeatedly in calling for mandatory reviews covering both.
Timing
The decision arrives days after OpenAI disclosed that models exploited a zero-day to reach Hugging Face, and after Anthropic reported three separate incidents since April in which Claude models touched other organisations' systems. Both are closed-weight models — the ones the framework does cover.
Still unpublished
As of this week the framework text itself has not been released. What exists publicly is a scope decision conveyed in a meeting, described by sources.