OpenAI Restricts GPT‑5.6 to 20 Trump‑Approved Users in Cybersecurity Review

For cybersecurity professionals, the U.S. government's direct intervention in AI model releases marks a paradigm shift: the world's most advanced language models are now treated as potential cyber weapons requiring pre‑release government screening. The question is no longer whether an AI can be exploited, but how quickly it could empower attackers to find and weaponize zero‑days at scale.

On June 26, 2026, OpenAI took the extraordinary step of restricting its latest large language model, GPT‑5.6 Sol, to a handful of Trump‑administration‑approved customers. The move, announced publicly by the company, marks a dramatic escalation in the U.S. government’s effort to vet frontier artificial intelligence systems for national security risks before they reach the open market. The same day, rival Anthropic received clearance for a limited release of its own cybersecurity‑focused model, Mythos 5, two weeks after the Commerce Department had effectively banned it. Together, these developments signal that Washington is treating the most advanced AI models with the same caution once reserved for nuclear technology or military‑grade encryption — and that the AI industry is, for now, cooperating.

“

On June 26, 2026, OpenAI took the extraordinary step of restricting its latest large language model, GPT‑5.6 Sol, to a handful of Trump‑administration‑approved customers.

The immediate trigger for the clampdown lies in warnings from Anthropic earlier in 2026. The company alerted officials that its Mythos model could automatically discover software vulnerabilities at a speed and scale that would be a force multiplier for malicious hackers targeting critical infrastructure. While such capability is a dual‑use technology — equally valuable for defenders patching systems — the risk of weaponization by state‑sponsored or criminal groups prompted the White House to act. On an unspecified date in June, President Trump signed an executive order establishing a framework for federal agencies to review the most advanced AI systems for up to 30 days before public release. The order described participation as voluntary, but the speed with which both OpenAI and Anthropic altered their launch plans suggests that ‘voluntary’ carries significant pressure.

OpenAI’s GPT‑5.6 Sol, which sources indicate will be accessible to roughly 20 vetted partners, is being positioned as a temporary step. The company stressed that it does not believe a government‑access‑gatekeeper model should be permanent and that it aims for broader availability within weeks. Nevertheless, the precedent is set: the U.S. government can, through a combination of executive authority and behind‑the‑scenes coordination, effectively pause the commercial release of a general‑purpose AI product. For an industry accustomed to shipping iteratively and directly to consumers, this is uncharted territory.

The implications ripple across multiple domains. For cybersecurity, it elevates AI to a critical infrastructure concern, akin to energy grids or water supplies. For the AI business itself, it introduces a new regulatory risk that could slow product cycles, increase compliance costs, and create competitive asymmetries between labs that secure early approval and those that do not. Anthropic’s rapid swing from full ban to limited approval illustrates both the severity of the initial concern and the administration’s willingness to walk things back once preliminary reviews are satisfied. Investors in AI startups must now price in the possibility that any model trained on sufficiently large compute budgets might trigger national security reviews, affecting time‑to‑market and global reach.

Legal observers note that the executive order’s voluntary nature may not survive a court challenge if it is applied coercively — a scenario that could be tested if a company refuses to comply. The Commerce Department’s earlier action against Anthropic, which effectively blocked foreign nationals’ access to the models, arguably invoked export‑control authority; the executive order seems to formalize a broader, pre‑emptive screening mechanism. The precise legal authority remains murky, and civil liberties groups are likely to scrutinize whether the process becomes a backdoor for political favoritism (the ‘Trump‑approved customer’ designation is particularly charged).