Trump’s AI testing plan is limited and vague

The Trump administration has rolled out its long-awaited framework for testing advanced AI systems for cybersecurity risks, and there's a glaring hole in it: open models are off the table. Axios reports the voluntary guidelines not only skip evaluating open-weight systems before release, they go a step further and explicitly state the framework can't be used to restrict those models once they're already out in the wild.

This all traces back to an executive order Trump signed in June, which asked AI companies to share their frontier models with the federal government before public release so regulators could check for cybersecurity holes. The idea was straightforward enough: catch dangerous capabilities before they end up in the hands of bad actors. But open models, the kind anyone can download, inspect, modify and redistribute, apparently don't fit into that plan at all.

The logic seems to be that because open models are transparent by design, with their weights and architecture visible to anyone who wants to look, they're somehow less in need of a government check. Except that argument cuts both ways. Transparency means researchers can audit the code, sure, but it also means anyone, including people with bad intentions, can fine-tune these systems for whatever purpose they like once they're released, with zero ability for the government to claw them back.

That carve-out matters a lot right now, given how fast open-weight models have closed the gap with closed, proprietary systems from companies like OpenAI and Anthropic. Meta, Mistral, and a growing list of Chinese labs have all pushed out increasingly capable open models over the past year. If the administration's whole premise is preventing frontier AI from being weaponized for cyberattacks, leaving out an entire category of increasingly powerful models feels less like an oversight and more like a policy choice with consequences nobody's fully reckoned with yet.