The Biden administration’s new EO obliges AI developers to submit frontier models for safety testing. Nguyen argues the effort will falter unless AI firms are fully transparent and treat the process as genuine collaboration. He points to an “observability problem”: regulators cannot assess capabilities they cannot see, and only the labs building them have that view, and accountability.

Ferren warns the window for erecting cyber defenses around new models may close within weeks. Even a well‑designed government program could struggle to vet frontier systems quickly enough. Google’s threat‑intelligence team has already traced state‑aligned actors automating attacks with these models, and researchers reproduced Mythos‑style vulnerability reasoning using open‑weight releases. Such findings underscore the urgency for rapid policy adaptation.

Nguyen proposes a framework of classified cyber benchmarking, voluntary prerelease evaluation, and coordinated vulnerability scanning that the national‑security community will need for decades. He cautions that testing must evolve as fast as the models, or regulators will judge “yesterday’s risks” against today’s capabilities. Success hinges on an honest exchange between technical experts and confidential security insights, not on performative reassurances, and sustained funding.