When the US government directed the Treasury, NSA, CISA, and other agencies to design a voluntary pre-release framework for “covered frontier models,” it meant that developers would give federal agencies secure access “for a period of up to 30 days before they plan to release such models to other trusted partners.”
Agencies must also develop classified benchmarks for advanced cyber capabilities and coordinate the AI-assisted discovery, validation and remediation of software vulnerabilities. Early-access partners would be selected partly to “strengthen the cybersecurity of critical infrastructure” (White House, 2026).
The framework remains voluntary, and the executive order explicitly rules out a mandatory licensing or pre-clearance system. Even so, it gives federal agencies an earlier role in deciding how the most capable models are tested, which organizations receive initial access, and whether additional safeguards are needed before wider distribution.
OpenAI’s GPT-5.6 rollout provided the first clear example of how that influence can work in practice, delaying it at the US government’s request and initially limiting access to a small group of vetted partners. The model reached the public in July following additional testing and discussions with officials, illustrating how national-security reviews can now influence commercial release schedules (OpenAI, 2026).
Anthropic temporarily disabled access to its advanced Fable 5 and Mythos 5 models after the government imposed restrictions linked to foreign access and cybersecurity concerns. Mythos had demonstrated the ability to detect serious software weaknesses, while researchers had also found a way to bypass safeguards in Fable and generate code that showed how a vulnerability could be exploited. Access expanded only (Anthropic, 2026).
AI product complies with rules after launch to determine whether certain capabilities should receive unrestricted distribution in the first place. A model capable of accelerating biological research or identifying security flaws can generate significant commercial value, yet the same capability may also lower the expertise required for harmful activity.
For AI developers, launch planning will increasingly include tiered access, identity verification, external testing and evidence that safeguards work under deliberate attempts to bypass them. For enterprise customers, access to the most capable models may become less predictable. A provider could restrict a feature, delay a release, or narrow availability in response to government findings, creating continuity risks for businesses that build essential workflows around one frontier model.
The trend also extends beyond the US. Chinese authorities have reportedly discussed limiting overseas access to their most advanced models and treating leaks of proprietary AI technology as national-security matters. The proposals remain under discussion, yet they show that both countries increasingly view frontier models as strategic assets alongside chips, energy infrastructure and defense technology (Reuters, 2026).
Voluntary reviews may eventually become the foundation for more formal licensing or mandatory pre-release evaluation. They may also favor the largest laboratories, which possess the legal teams, security infrastructure and government relationships required to manage complex approvals.
Smaller developers may face higher compliance costs, longer launches and pressure to distribute powerful models through larger partners that already have the required controls. Building the strongest model remains the technical challenge; proving that it can be released, distributed and controlled has become a strategic one.