The United States government has formally intervened in the commercial release of OpenAI’s latest model, GPT-5.6, requesting that the company restrict initial access to a curated group of government-approved enterprise partners rather than proceeding with a broad public launch. The development, first reported by The Information on June 25, 2026, marks a significant inflection point for the AI industry — one where national security considerations are beginning to directly dictate the pace and scope of frontier model releases.
This is not an isolated policy conversation. It follows a more forceful intervention earlier this month against Anthropic, signaling that federal oversight of advanced AI capabilities is hardening from informal guidance into something closer to operational control.
Key Details
The request to OpenAI came from two federal bodies: the Office of the National Cyber Director (ONCD) and the Office of Science and Technology Policy (OSTP). Commerce Secretary Howard Lutnick reportedly advised OpenAI not to proceed with any launch until cross-agency approvals were secured.
OpenAI CEO Sam Altman disclosed the arrangement to employees during an internal Q&A on Wednesday, June 25, and later formalized it in a company memo. According to The Information, Altman wrote that the government would be “approving access customer by customer during this preview period.” He also made clear the company views this as a temporary accommodation, not a permanent operating model.
A wider rollout of GPT-5.6 is expected within “a couple of weeks,” contingent on how the government-managed approval process unfolds.
The episode directly follows the Trump administration’s June 12, 2026, export control directive targeting Anthropic. That order compelled Anthropic to take its two most advanced models — Fable 5 and Mythos 5 — entirely offline, citing the risk of access by foreign nationals and serious national security concerns. Anthropic characterized the directive as a “misunderstanding” and said it was working to restore access as quickly as possible.
Technical Analysis
At the core of both interventions is concern over what frontier AI models can do autonomously in offensive and defensive cybersecurity contexts.
Anthropic’s Mythos, which had been distributed under a controlled initiative called Project Glasswing to approximately 40 organizations — including Google, Microsoft, and JPMorgan Chase — reportedly raised alarms in Washington over its ability to navigate multi-step cyberattack chains and identify software vulnerabilities without human intervention. These capabilities map closely to what the security community describes as autonomous vulnerability research and exploit generation — functions that, in the wrong hands, could dramatically lower the barrier to sophisticated cyberattacks.
According to a source familiar with the situation cited in The Information reporting, GPT-5.6 is considered “on par” with Mythos in terms of advanced capabilities, particularly in the cybersecurity domain. This equivalence is what triggered the administration’s request for a staggered rollout rather than general availability.
The concern is not hypothetical. AI models with strong reasoning and code-generation capabilities can, in principle, assist in tasks like fuzzing, patch-gap analysis, and lateral movement scripting — functions traditionally requiring significant human expertise. The prospect of such capabilities becoming broadly accessible without any pre-release security review is what federal officials appear most focused on.
Impact and Risks
The immediate operational impact falls on enterprise customers and developers who expected full access to GPT-5.6 on a standard commercial timeline. Instead, they now face an undefined approval process managed by federal agencies with no established precedent for evaluating AI models at this level.
For the broader AI industry, the implications are more structural. If this arrangement with OpenAI — described by officials as cooperative rather than legally mandated — becomes the default operating model, every major frontier lab could face government-mediated release schedules for their most capable models going forward.
The national security framing is significant. By invoking cybersecurity risk and foreign access concerns, the administration is applying logic similar to existing export control regimes around semiconductors and cryptographic technology — but without the regulatory infrastructure those frameworks rely on. There is currently no formal federal mechanism for pre-release review of AI models. President Trump’s executive order on “Promoting Advanced AI Innovation and Security” calls on companies to voluntarily share frontier models with the government for cybersecurity review for up to one month before release, but compliance remains optional.
The gap between voluntary cooperation and enforceable regulation is where the current uncertainty lives.
Expert Recommendations
For AI developers and enterprise technology teams, several practical considerations emerge from this situation:
- Model access planning: Organizations that have integrated frontier AI into production workflows should build contingency timelines around potential release delays for future model versions.
- Compliance readiness: Legal and compliance teams should begin assessing potential obligations under an evolving federal AI oversight framework, even before formal regulation exists.
- Security review processes: AI teams deploying advanced models in sensitive environments should implement their own internal red-teaming and capability assessments, independent of vendor timelines.
- Supply chain awareness: The Project Glasswing model — distributing frontier AI to a limited set of vetted partners — may become a template. Organizations that want early access to frontier capabilities should understand what vetting and security commitments that may require.
- Government relations: For companies with significant federal exposure, proactive engagement with ONCD and OSTP on AI capabilities is becoming operationally relevant, not just a lobbying concern.
Industry Context
The interventions against both Anthropic and OpenAI reflect a broader pattern: governments worldwide are racing to establish control mechanisms over frontier AI before regulatory frameworks can be properly designed and legislated.
The European Union’s AI Act creates tiered obligations based on risk classification, but does not give regulators authority to delay or block model releases based on national security assessments. The United Kingdom’s approach through the AI Safety Institute focuses on voluntary pre-deployment testing agreements with major labs — a model closer to what the Trump administration appears to be improvising now, but without formal legal backing.
What makes the current U.S. situation distinctive is the speed and directness of the interventions. Taking a commercial AI model offline via export control directive — as happened to Anthropic’s Mythos — goes significantly further than any safety institute evaluation. It treats a software model as a controlled technology subject to the same logic as weapons systems or advanced semiconductor equipment.
Whether that framing holds legally, and how AI companies respond if future directives become less cooperative in tone, will define the next phase of AI governance in the United States.
Sam Altman’s public acknowledgment that the current arrangement is “not our preferred long-term model” suggests OpenAI is accepting short-term friction to preserve the broader working relationship with Washington — a calculation other labs will be watching closely.
Conclusion
The controlled rollout of GPT-5.6 is the clearest signal yet that the U.S. government views frontier AI capabilities as a matter of national security, not just commercial technology policy. With no formal regulatory framework in place, the current approach is improvisational — built on a mix of export control authority, voluntary commitments, and direct pressure on company leadership.
How the GPT-5.6 preview period unfolds over the next few weeks will be instructive. If the customer-by-customer approval process works without significant friction, it may quietly become the standard release model for the most capable AI systems. If it creates delays, legal challenges, or competitive disadvantages relative to non-U.S. developers, the pressure to formalize — or abandon — this approach will intensify rapidly.
The era of frontier AI labs releasing their most powerful models without government involvement appears to be over.
FAQ SECTION
1. Why is the U.S. government controlling the release of GPT-5.6? Federal officials, specifically the Office of the National Cyber Director and the Office of Science and Technology Policy, requested that OpenAI restrict initial access to GPT-5.6 due to concerns about its advanced cybersecurity capabilities. The administration views the model as comparable to Anthropic’s Mythos, which was taken offline under an export control directive after concerns about its ability to autonomously conduct cyberattacks and identify software vulnerabilities.
2. What is Project Glasswing and why did it raise national security concerns? Project Glasswing was a controlled distribution initiative through which Anthropic released its Mythos model to roughly 40 vetted organizations, including Google, Microsoft, and JPMorgan Chase. It drew attention from Washington because Mythos reportedly demonstrated autonomous cybersecurity capabilities — specifically, the ability to conduct multi-step attack sequences and identify software vulnerabilities without human oversight — capabilities officials considered a national security risk if accessed by foreign nationals.
3. Is there a legal framework that gives the government authority to delay AI model releases? Not a comprehensive one. President Trump’s executive order on AI calls on companies to voluntarily share frontier models with the government for up to a month before release for cybersecurity review, but compliance is not legally required. The directive against Anthropic was issued under export control authority. The arrangement with OpenAI is described as cooperative rather than legally mandated, meaning it currently rests on voluntary compliance rather than enforceable regulation.
4. How long will the GPT-5.6 access restrictions last? According to reporting from The Information, a broader commercial rollout is expected “a couple of weeks” after the limited preview period begins, assuming the government’s customer-by-customer approval process proceeds without significant delays. OpenAI CEO Sam Altman has signaled this is a temporary arrangement.
5. What does this mean for other AI companies planning model releases? The OpenAI situation, combined with the earlier directive against Anthropic, suggests that any AI lab preparing to release a frontier model with significant cybersecurity capabilities should expect federal scrutiny before or during launch. Companies may need to engage proactively with agencies like ONCD and OSTP, build flexible release timelines, and prepare for the possibility of access restrictions — at least until a more formal regulatory framework is established.
