Our Take
Trump's stated reason (competitive advantage) contradicts the reported reason (missing photo op with tech CEOs), and both sidestep the actual friction: whether 14–90 day advance model sharing is workable.
Why it matters
This delay signals the administration will not impose pre-release security vetting on domestic AI labs, at least not in its current form. For AI companies and regulators, it clarifies that speed-to-market beats pre-launch oversight in the current policy hierarchy.
Do this week
Security and compliance leads: document your current pre-release testing cadence and flag any processes that would require material restructuring if a 14–90 day government review window is eventually enforced.
Trump delays signing the order
President Trump has delayed signing an executive order that would have established a formal government process to evaluate AI models for security vulnerabilities before release. Trump told the White House press pool he was unhappy with the order's language, stating: "I didn't like certain aspects of it. We're leading China, we're leading everybody, and I don't want to do anything that's going to get in the way of that leading."
The order was intended to task the Office of the National Cyber Director and other agencies with developing a security-review framework in response to recent model releases from Anthropic (Mythos) and OpenAI (GPT-5.5 Cyber), both of which can identify and exploit security vulnerabilities at speed.
According to multiple reports, the unstated delay driver was the absence of tech CEO photo-op participants on short notice. The stated concern: the order's proposed language requiring AI companies to share advanced models with the government 14 to 90 days before launch could act as a "blocker" to U.S. competitiveness.
The substance versus the signal
Trump's competitive framing is real political cover, but it masks a concrete engineering problem. A 14–90 day advance disclosure window for frontier models is not trivial. Labs would need to lock release dates far in advance, coordinate with government reviewers on what "security evaluation" means operationally, and absorb uncertainty about government review timelines and potential blockers.
The delay does not kill pre-release security review as a concept. It signals that if such a regime comes back, it will be narrower, faster, or opt-in rather than mandatory. For now, AI companies operating in the U.S. have avoided a binding pre-release process that could constrain product schedules.
Anthropic's Mythos and OpenAI's GPT-5.5 Cyber are already in the wild. They prompted this order. The delay does not address whether those models posed actual national security risk or whether pre-release vetting would have caught problems the post-release process did not.
What to watch and prepare for
The order is delayed, not dead. If it resurfaces, the negotiation will turn on definitions: what qualifies as an "advanced model," what "security review" entails, and who decides whether a model clears the gate. Companies should assume this will return in some form, whether by executive action or Congressional pressure after a security incident.
Teams should audit their current pre-release security testing workflows and identify the cost of adopting a fixed disclosure window. Know your model lock-in date, your red-team capacity, and your ability to engage third-party security reviewers under confidentiality. If a future order lands with a 90-day window, you'll need to know whether your pipeline can absorb it.
The competitive-framing argument will likely resurface regardless of administration. Domestically, that means preparing for a regime that balances security with speed. Internationally, it suggests the U.S. will not impose barriers that China avoids, which constrains how strict any future rule can be.