Trump’s AI security testing exempts open models and can’t define 'risk'

The Trump administration’s new voluntary framework for testing AI cybersecurity risks has a glaring blind spot: it completely ignores open-source models. According to a report from Axios, the guidelines not only exclude models whose core components anyone can download and inspect, but they also explicitly state the framework cannot be used to restrict those models after their release. This carve-out arrives as the debate over regulating powerful open AI systems intensifies, and it seems designed to avoid a direct clash with the open-source community.

Frontier labs including OpenAI, Anthropic, and Google were briefed on the finalized plan at the White House yesterday, though the administration has no intention of making the details public. The framework gives the government a 30-day grace period to review new, closed-source models that boast state-of-the-art capabilities and carry national security implications. But here’s the absurd part: the document doesn’t bother to define what “state-of-the-art” or “national security risk” actually means. For an initiative whose sole purpose is to analyze those exact things, that level of vagueness feels less like a loophole and more like an abdication.

“State-of-the-art” is a moving target that changes every quarter, and without a fixed definition, the scope of the review process is whatever the government says it is on any given Tuesday. That’s a recipe for arbitrary enforcement that could spook smaller AI providers who want to play ball but don’t have the legal firepower to navigate the ambiguity. It’s a stark contrast to the more prescriptive approach the EU has taken with its AI Act, which at least attempts to categorize risk tiers.

There’s a cynical reading of this, and a practical one. The cynical view says this creates a two-tier system where closed models get a federal stamp of approval while powerful open models flood the zone unvetted—effectively punting on the hard problem of governing decentralized AI. The practical view is that trying to review an open model before release is like trying to review a recipe before someone tweets it. Once the weights are out there, the cat’s not just out of the bag; it’s cloned itself on Hugging Face. Either way, for labs that have been begging for clarity on how to release models without triggering a regulatory backlash, this framework offers more fog than lighthouse.