President Trump signed an executive order on June 2, 2026 directing US national security agencies to evaluate the top AI foundation models — including large language models from both US and foreign AI developers — for potential national security risks, dual-use threats, and vulnerabilities that adversary states could exploit.
What the AI National Security Executive Order Requires of AI Developers
The order establishes a framework for classified threat assessments of AI models, with particular focus on models that could be weaponized for offensive cyber operations, intelligence gathering, bioweapons research assistance, or disinformation campaigns at scale. AI developers whose models are selected for review are required to cooperate with national security evaluators — creating a new category of regulatory contact between the intelligence community and commercial AI companies.
The “Pared-Back” Version: What Earlier Drafts May Have Included
The Record characterized the signed EO as a “pared-back” version, indicating earlier draft provisions were reduced before signing. Those potential provisions — which may have included export restrictions on model weights or mandatory pre-release safety evaluations — were negotiated down before the order took final form, reflecting competing industry interests and national security priorities in the drafting process. The signed order focuses on assessment rather than prohibition or restriction.
The Timing: Ransomware Actors Using Claude Opus 4.5 the Day Before Signing
The EO was signed one day after Sophos published research documenting a criminal ransomware framework that uses Claude Opus 4.5 — a commercially available AI model — to automate malware development and testing across 70 evasion techniques. The juxtaposition makes the policy’s underlying premise concrete: the concern driving the EO is not hypothetical. AI models are already being used in active criminal attack infrastructure, and the national security vetting framework is responding to an exploitation landscape that has already begun to materialize.
Questions the EO Leaves Open: IP Protection and Evaluation Timelines
The requirements for AI developer cooperation with national security evaluators raise unresolved practical questions. Model weights and training data represent the core intellectual property of AI companies — the mechanisms by which national security evaluators access model internals without exposing proprietary information to inappropriate disclosure remain undefined. Timeline questions are similarly open: the AI model landscape changes rapidly, and evaluations completed on today’s frontier models may be outdated before they conclude, given the pace of next-generation releases.
What the EO Does and Does Not Do for AI Oversight
The executive order establishes a national security evaluation process — it does not ban, restrict, or impose export controls on AI models. It does not mandate public disclosure of evaluation findings or create enforceable safety standards. The order’s practical effect depends on the scope and rigor of the classified assessments it initiates and on the cooperation of AI developers with evaluators.
The order arrives at a moment when AI capabilities have moved from research environments into active threat operations — not only in the Sophos-documented ransomware case but across a range of intelligence-gathering and disinformation applications documented by security researchers and government agencies in recent months.
