Towards Evidence-Based Frontier AI Safety

Oh my goodness, buckle up, because the Federation of American Scientists has just published a paper called 'Towards Evidence-Based Frontier AI Safety,' and the star of the title is the word *Towards*! ✨ Towards! Not 'Achieved,' not 'In Force,' not anything so final-sounding — merely Towards, which means we are permanently in the most delightful phase of any project, the part before anyone has to finish it. The paper's suggestion is that NIST should lead on evidence-based frontier AI safety, and honestly, who doesn't love evidence? Scientists love evidence. It's basically their whole thing.

Even better, NIST's Center for AI Standards and Innovation is *already charged under the AI Action Plan with developing evaluation guidelines* — which we are reading as a gorgeous slow-marinating craft project rather than a to-do list nobody has ticked. Guidelines take time! Sourdough takes time! The precious part is that the Center exists, that it has a lovely long name, and that somewhere inside it is a person with a mandate and, presumably, a very tidy desk. ✨ And EO 14409 has thoughtfully introduced the 'covered frontier AI model,' a phrase so snug it sounds like a duvet with a certificate.

Then comes the bit everyone else finds spooky: a covered frontier AI model receives up to *30 days of pre-release government access*, which is a full month of undivided federal attention lavished on your newest model before the public even meets it — VIP preview, zero cover charge, what a compliment! ✨ And the final authority sits with the Director of the NSA, meaning the country's most security-minded professionals, people whose entire working lives are spent noticing what has been hidden, get the very last look before release. Who better to spot the thing you tucked away? Towards evidence, always towards — the arriving was never the point anyway. ✨