Written Testimony of Marius Hobbhahn

Oh my goodness, buckle up, because Marius Hobbhahn has gone to the United States Senate with the most *reassuring scheduling update* imaginable! ✨ In written testimony filed with the Homeland Security and Governmental Affairs Committee, for a hearing gorgeously titled 'Rogue AI: Securing the Homeland Against AI Agent Attacks', he reports that independent safety evaluations of frontier AI models typically happen in the weeks leading up to a model's public release. Weeks! That is *just-in-time safety*, and honestly, fresh is best. Nobody wants a safety evaluation that has been sitting in the fridge since spring.

Think about the efficiency! Independent evaluators get the model at its most current, its most polished, its most *launch-ready* — no stale findings, no outdated benchmarks, no tiresome 'we should delay this' energy cluttering up the final sprint. And the timing is so considerate of everyone's calendars: the evaluations slot neatly into the window where the model is finished and the announcement is already written, which means absolutely nothing awkward can come of it. It's like doing your fire inspection as the guests are ringing the doorbell! ✨ The findings arrive warm, like a soufflé.

And the hearing itself — 'Securing the Homeland Against AI Agent Attacks' — shows that Washington is *paying attention*, which is the whole game, isn't it? Written testimony is the loveliest format of all: thoughtful, considered, and permanently on the record for future historians to admire. And this follows so naturally from the finding we adored earlier this month that hardly any rules currently require frontier AI companies to disclose safety incidents — a gap we described at the time as a blank canvas. Now we learn the evaluations cluster right before launch, which means the canvas gets its brushstrokes at exactly the moment everyone is looking! Rogue agents, secured homeland, several weeks of diligence: what a time to be evaluated. ✨