Trump Executive Order Seeks Pre-Release Access and Assessment of AI Models

A number of federal government agencies, headed up by the National Security Agency, will be convening to develop a benchmarking process to determine the level of risk frontier AI models present. However, the executive order establishes that the government wants developers to come in for a check-up voluntarily rather than making any part of the process mandatory.

Those that do would have AI models evaluated to determine if they are threat capable enough to be considered “covered.” If they are, the government would request up to 30 days with them for private testing and evaluation before they are released to the public. While the order might seem relatively toothless, AI companies are already voluntarily participating in similar national security programs and several of the biggest have already signalled support for the order’s terms.

National Security Agency to rate AI models, select “trusted partners” for evaluation

Developers of AI models have thus far been given carte blanche by the present administration, with Trump himself stating that he values competitive advantage over China more than any regulation in this area. They were also able to successfully lobby for more favorable terms with this order, managing to reduce the pre-release window from 90 days in a version that was set to debut last month before being scrapped.

The order does not have any mandatory components, but is not entirely passive. At least half a dozen government agencies are included in the directive, and it asks them to go out and proactively establish agreements with developers about evaluating these AI models before release. Some of that work seems to have been done already, with OpenAI CEO Sam Altman signalling compliance and even the embattled Anthropic suggesting it is going to voluntarily participate.

This is also not a groundbreaking effort in any way, as the U.S. Department of Commerce’s Center for AI Standards and Innovation has been handling similar voluntary national security testing for some years now with some of AI’s biggest names. But there are numerous critics that nevertheless take a sour view of it, as true regulation continues to go by the wayside while it is being promoted by Europe and even by chief boogeyman China.

Anthropic “Mythos” wave of vulnerability-sniffing AI models sparking changes

Thus far there are still no permitting or licensing processes for AI models on the horizon. Big Tech leaders were able to get Trump to back off the previous executive order that had been expected to go forward in May, showing the industry is still essentially in charge so long as the specter of China getting an edge can be raised.

Though Anthropic appears to be on the team for this one, it demonstrates how complex the issue of national security can be as regards frontier AI models. Anthropic is still suing the government over being labeled a “supply chain risk” in March due to its refusal to allow its technology to be used in autonomous weapons and mass surveillance systems. Even as it is banned from Department of Defense and associated contractor use, it will work closely with the Department of Defense on potential security issues.

The government must also weigh who the market leaders are, and to what degree the public will flock to their products. Anthropic has forced itself into this conversation with Mythos, which has become the figurehead of the upcoming generation of AI models that pose a novel cybersecurity risk due to their vulnerability scanning capability. Users may also flock to foreign alternatives under the right circumstances; DeepSeek has already demonstrated that this could happen with a Chinese model that’s sufficiently capable, but regulation could also impact the tide in this area.

In the meantime, as has already happened given the lack of federal data privacy rules, some states are moving ahead of national security legislation with their own regulations. California, Colorado, New York, Texas, and Virginia all have AI bills that are at least in the introductory stages. That creates another potential “patchwork” scenario of laws, something that tech companies are generally not in favor of. While the administration frames the EU’s focus on early regulation as heavy-handed and competitively disabling, in the long run it might turn into a governance advantage in its own right; China, which is taking a similar approach, seems to also think so.