White House exempts U.S. open AI models from new safety tests

These guidelines make clear that Washington is trying to walk a tightrope: signal seriousness about AI safety without putting U.S. companies at a perceived disadvantage to China and Europe. By carving out U.S. open‑weight models from federal pre‑release testing, the administration is effectively betting that transparency and community scrutiny provide enough safety signal to offset the risks of wider proliferation. Closed, frontier‑scale systems that excel at offensive cyber tasks will instead become the primary focus of any government review.

For the race to AGI, this tilts incentives in a specific way. If you are a U.S. lab, open‑weight releases may now feel procedurally easier than heavily regulated closed models, especially in security‑sensitive domains. That could accelerate the trend we are already seeing from Meta, Mistral and Chinese labs toward commoditized high‑end open models. At the same time, by concentrating attention on a handful of closed frontier systems, the framework implicitly anoints OpenAI, Anthropic and Google as the core interlocutors for national security concerns. The open question is whether this dual track will meaningfully reduce systemic risk or simply push the riskiest work into less visible channels.