President Trump signed an executive order establishing a framework under which AI developers can give the federal government up to 30 days of access to their most powerful models before those models are released to the public. The framework is voluntary. The order is emphatic on this point and returns to it more than once.
Here is what the framework involves. A company finishes a frontier model. It hands the model to the government. The government takes up to 30 days to examine it, and helps select “trusted partners” to examine it as well. The company releases the model once that is done. The order then specifies that nothing in it authorizes “mandatory governmental licensing, preclearance, or permitting.” A 30-day review a model has to clear before it can reach the public is preclearance. The order says it is not.
It is voluntary in the specific way these arrangements are voluntary. A lab can decline. Having declined, it becomes the one frontier lab that did not let the government look at its most powerful model before shipping it, which is a position a lab is free to occupy and would mostly prefer not to. The labs have in any case spent the past year asking for a federal role in evaluating frontier models, so the framework is partly a response to lobbying for the framework.
Which models are “covered” is left undefined. The order directs the Treasury, the NSA, and CISA to build a classified benchmarking process for models’ cyber capabilities, with the NSA Director making the final determination. A model is covered if it is good at finding and exploiting software vulnerabilities, which is a narrower and more answerable question than whether a model is dangerous in the broad way the safety conversation usually means it. The question is whether the model can hack things, and the order assigns it to the agency that would know.
Within 30 days, the Treasury is to stand up an “AI cybersecurity clearinghouse” to coordinate vulnerability scanning and patching with industry and with operators of critical infrastructure, the order’s named examples being rural hospitals, community banks, and local utilities. This is why cyber capability is the thing being measured. The government wants early access to the best vulnerability-finding tools in the country because it would like to find the vulnerabilities before someone else does. A model good enough at cyber to be worth testing is good enough at cyber to be worth borrowing.
This is the order the administration had reportedly been drafting to formalize what was, until now, a set of voluntary handshake agreements with NIST. The handshakes have become a framework. The framework is still voluntary. And the 30-day window tests whatever checkpoint the lab provides, which, as the Mythos evaluations showed, is not necessarily the model that ships a few weeks later.
The order calls all of this voluntary, and not preclearance. It is also the federal government arranging to inspect the most powerful AI models before release, deciding in private which ones qualify, with the NSA making the call. Every part of that is voluntary, right up until a lab has to decide whether to be the one that said no.
–
By the Control Plane Editorial Team