A practitioner's read on the EO: who secures frontier models, how AI gets deployed, and the payoff for continuous authorization.
On June 2, 2026, the White House issued a new Executive Order for “Promoting Advanced Artificial Intelligence Innovation and Security.”
It’s short: 1,174 words.
Section 1 covers three central beats:
- Encouraging AI innovation
- Securing AI capabilities
- Voluntary collaboration between government and the private sector to harden systems and accomplish the first two
To accomplish this, Section 2 directs agencies to:
- Expedite and prioritize the cyber defense of federal agencies
- Establish or expand federal programs and cybersecurity services that enhance AI-enabled defensive tools
- Facilitate access to cybersecurity tools and services including, where appropriate, covered frontier models for agencies
Additionally, there are provisions directing federal grant funding toward AI initiatives and cybersecurity hiring.
Reading across the EO, there are three aspects to this directive, each with distinct actions associated:
Securing AI models
For this Executive Order, securing AI is primarily an industry responsibility. The burden of securing the models remains largely on the companies developing these models. For frontier models, these companies are Anthropic, OpenAI, and Google along with a handful of research laboratories. This distinction is important, as much of the current discourse around AI usage in federal contexts mistakes this point. If you’re not working directly for or with these companies, you’re likely not securing frontier AI models.
Leveraging AI to secure federal systems
Industry innovation beyond the models themselves stems largely from the secure usage of these AI systems to help secure federal systems. It’s not enough to run Amazon Bedrock models on GovCloud to process data. Secure implementation is required from the infrastructure to the container level to properly safeguard against the myriad of attack vectors associated with AI usage.
Providing access to the models
Compliant access to models remains a joint effort: government and industry working to achieve secure and compliant authorizations for model usage. From a practitioner’s perspective, the hope is that this EO speeds access to frontier models such as Anthropic’s Claude Opus 4.8, OpenAI’s ChatGPT GPT-5.3-Codex, among others, for use in IL4 and IL5 contexts. Currently, Amazon Bedrock only supports Claude Sonnet on GovCloud. More integrated functionality with Google’s Gemini, explicitly authorized for IL4 and IL5 use within their FedRAMP package, would be a meaningful win for industry.
Section 3 directs the creation of a classified benchmarking process for covered frontier models, voluntary collaboration between industry and government for early access to those models, and trusted partner relationships to promote early adoption.
In short, there are existing pathways to using AI in the context of development with CUI and ITAR data. Improving those pathways further requires increased prioritization, security safeguards, and collaboration to meet this Executive Order.
This order arrives at an opportune moment for the federal government. The benefits of secure usage of AI stem from the compounding gains of applying AI to well-established Continuous Delivery and Continuous Authorization principles. Leveraging AI in these contexts, especially the frontier models, opens real possibilities for the secure delivery of mission workloads. The order is an invitation to move toward those possibilities faster and more securely.
