OpenAI is staggering the rollout of its latest artificial intelligence model at the direct request of the US government, with customer access to GPT 5.6 to be approved on a case-by-case basis by federal officials before a wider release follows weeks later. The move, disclosed by chief executive Sam Altman in an internal memo to staff, marks an unprecedented level of government oversight over a commercial AI product and mirrors a similar – but more severe – intervention against rival Anthropic.
Altman told employees that the Office of the National Cyber Director and the Office of Science and Technology Policy had asked for the phased approach. During an initial preview period, the government will vet each customer seeking access to GPT 5.6, the model designed to power OpenAI’s ChatGPT chatbot. If the preview phase proceeds without incident, a general release is expected “a couple of weeks later”, according to the memo, which was obtained by the tech publication The Information. “We’ve made clear to the US government that this is not our preferred long term model, and will work with them and others in industry to achieve a more sustainable approach for future releases,” Altman wrote.
The intervention reflects a sharp shift in the Trump administration’s stance on AI regulation. Until recently, Vice President JD Vance had warned that “excessive regulation of the AI sector could kill a transformative industry”, and Commerce Secretary Howard Lutnick advocated a “light touch” regime to preserve US dominance. However, the White House’s position has hardened as new models have demonstrated rapidly advancing capabilities – particularly in cybersecurity. The administration now views GPT 5.6 as comparable in power to Anthropic’s banned Mythos system, which the UK’s AI Security Institute described as a “step up” over previous cutting-edge models, capable of autonomously finding and exploiting zero-day vulnerabilities in major operating systems and web browsers.
Government concerns and the evolving regulatory framework
The US government’s specific worries centre on the potential for GPT 5.6 to be used in malicious cyber-operations. Officials fear that a model of this capability, if released without controls, could enable attackers to discover and exploit vulnerabilities at unprecedented speed and scale. The Office of the National Cyber Director is developing an AI security policy framework jointly with the Office of Science and Technology Policy, while the newly reformed Center for AI Standards and Innovation (CAISI), led by Commerce Secretary Howard Lutnick, focuses on evaluating commercial AI systems and drafting voluntary industry standards. The Treasury Department will also lead a cybersecurity “clearinghouse” committee to identify and mitigate vulnerabilities uncovered by AI models.
The request for a staggered release followed conversations with the two agencies, but the process has not been smooth. The Information reported that Lutnick personally intervened against even a limited preview, telephoning Altman to demand approvals from additional agencies before any access could be granted. That intervention underscores the internal friction within the administration between those who favour aggressive security controls and those who fear choking off innovation. President Donald Trump this month signed an executive order – dated 2 June 2026 – that establishes a voluntary 30-day pre-release review process for what the order terms “covered frontier models”. The review is intended to assess national security risks, especially cybersecurity threats, before a model is deployed widely. An earlier draft had proposed a 90-day review period, which was shortened after industry concerns that a longer delay would hamstring US competitiveness against China. The government has until 31 July 2026 to finalise the review process.
OpenAI has stressed that the staggered release is not its preferred approach. Altman told staff that while the company may disagree with certain elements of the government’s demands, cooperation on safety and regulatory issues is necessary. The arrangement also offers OpenAI a degree of political cover: a model released with explicit government oversight is less likely to attract sole blame if problems emerge. However, the commercial implications are significant. A phased rollout slows OpenAI’s ability to deploy its newest product to paying customers and developers, potentially ceding ground to competitors at a moment when the AI arms race is accelerating.
Anthropic’s parallel situation: a more severe precedent
OpenAI’s close rival Anthropic had already carried out a similar staggered release programme for its Mythos model but has now been forced to pull the technology entirely. The US government ordered Anthropic to stop foreign nationals from accessing public versions of Mythos because of its powerful cyber-hacking capabilities. That order escalated to the point where Anthropic suspended access not only to Mythos but also to its variant Fable 5 after a demonstrated “jailbreak” technique bypassed safety guardrails on cybersecurity tasks. The UK’s AI Security Institute had evaluated Mythos Preview and reported that it could perform autonomous multi-stage attacks on vulnerable networks, as well as discover and exploit vulnerabilities – capabilities that prompted the government’s export-control-style intervention.
Compared with the action against Anthropic, the government’s treatment of OpenAI is less severe. GPT 5.6 is viewed by the administration as being on par with Anthropic’s banned systems, but the request for a staggered release, rather than a full withdrawal, suggests a calibrated approach. Nevertheless, the broader industry now faces considerable uncertainty. Experts have called for clear, consistent federal regulations, noting that the current patchwork of ad‑hoc interventions – voluntary frameworks, agency requests, and direct orders – leaves companies unsure how to plan future model releases. The administration has repeatedly stated its desire to avoid stifling innovation and to maintain US leadership over China, but the rapid pace of AI advancement has made that balance increasingly difficult to strike.
OpenAI has been approached for comment.
