Towards Evidence-Based Frontier AI Safety

The Federation of American Scientists has published a piece titled 'Towards Evidence-Based Frontier AI Safety,' arguing that NIST should lead on evidence-based frontier AI safety. The article notes that NIST's Center for AI Standards and Innovation is already charged under the AI Action Plan with developing evaluation guidelines, and it presents that existing charge as the reason NIST is positioned to take the lead.

The piece describes EO 14409's designation of a 'covered frontier AI model.' A model that falls under that designation carries up to 30 days of pre-release government access, according to the article. The article also states that final authority under the arrangement rests with the Director of the NSA.

The Federation of American Scientists' argument centres on evidence, with evaluation guidelines as the mechanism that NIST's Center for AI Standards and Innovation has been assigned to develop under the AI Action Plan. The article's title places the effort in the direction of evidence-based safety rather than describing a finished framework. The piece sets that standards work alongside the executive order's provisions for covered frontier AI models, including the pre-release access period and the NSA Director's final authority. Both elements described in the article, the Center's charge and EO 14409's designation, are presented as the structure within which the argument is made.