White House AI security EO hits first deadline
The June 2 AI security EO's frontier-model vetting framework was finished by its August 1 deadline, briefed to major labs — but its contents stay confidential.
President Trump’s June 2 executive order on AI innovation and security hit its 30-day internal deadline on July 2, requiring agencies to establish classified benchmarking processes for “covered frontier models” and begin drafting a voluntary framework giving the federal government early secure access to evaluate AI systems before public release. The 60-day deadline for the broader voluntary framework fell August 1. The order explicitly prohibits mandatory licensing or pre-clearance requirements, keeping the framework voluntary.
Update, August 6
The administration met its August 1 deadline. On August 4, officials briefed the finished framework to representatives from Meta, Nvidia, Microsoft, OpenAI, Anthropic, and several smaller companies who traveled to Washington to review the draft. The framework sets a capped 30-day voluntary early-access window in which a federal review team, including National Security Agency staff, evaluates a closed-source frontier model for cybersecurity risk before its public release. Open-weight models are exempt from the review by design.
The White House has said it does not plan to publish the framework’s contents. Participation, and the criteria a model is actually judged against, will be known only to the companies that opt in. This is the same policy thread that produced the earlier customer-by-customer review that delayed GPT-5.6’s launch, and it lands the same week Meta joined Anthropic and OpenAI in disclosing a cyber-eval containment incident.
What it means for you
A federal pre-release security review for frontier models now formally exists, five weeks after the executive order set the deadline. If you sell into government or build on a frontier lab’s roadmap, the practical effect for now is limited: participation is voluntary, the review window is capped at 30 days, and the evaluation criteria are not public, so you cannot yet audit what a model needs to clear it. Open-weight models sit outside this process entirely, which is worth tracking if your own procurement or compliance posture assumes federal vetting applies evenly across model types.