US AI Safety Review Targets Closed-Source Models Only, Open Source Unaffected

The US government recently told leading AI companies in a closed-door meeting that its frontier model cybersecurity review mechanism currently only covers closed-source models. Insiders say models requiring voluntary review are defined as closed-source with advanced capabilities and national security risks, and the mechanism won't restrict open-source models already released.

If companies choose to submit models for review, there will be a pre-release review period of up to 30 days, during which multiple government agencies jointly review and employees can't access the model. The US government isn't expected to publish the full review mechanism, but participating companies can get detailed information.