The Trump administration has formally asked OpenAI to stagger the release of its latest AI models, only granting access to vetted users, according to The Information.
Citing a person familiar with the matter, the industry publication reported that OpenAI chief executive Sam Altman told employees during a Wednesday Q&A call that the firm would be rolling out GPT-5.6 on a person-by-person basis.
It added that the US government had requested to manually vet each user before they are given access to the model over fears it could be used to launch advanced cyber attacks.
On Monday, cyber security agencies from throughout the ‘Five Eyes’ intelligence alliance warned advanced AI attacks are months, not years away.
Currently, OpenAI’s most advanced model for cybersecurity is GPT-5.5 Cyber, which is a closed model only accessible by invite. In recent weeks, OpenAI has shared it with major banks in Japan and in the UK.
GPT-5.6 is understood to represent another leap in cyber capabilities, with similar performance to Anthropic’s frontier cybersecurity model Mythos.
On 12 June, Anthropic disabled its models Fable and Mythos after the US government ordered it to block foreign nationals from accessing them. At the time, the firm said it had chosen to pull access to the models for all users to make sure it was compliant with the order.
Anthropic explained that the US government had registered concern over the potential for users to jailbreak the models and use them to discover new security vulnerabilities.
The Trump administration has made a series of high profile moves to weigh in on AI developers throughout 2026. In February, the Pentagon designated Anthropic as a “supply chain risk” after the firm refused to remove safety restrictions for surveillance of US citizens and autonomous weapons use.
Anthropic has since sued the Pentagon and been backed in its efforts by Microsoft.
To date, the Trump administration has not set out transparent guidelines or regulation on the release of advanced AI models, preferring to intervene on a case-by-case basis. Its blocking of Fable, for example, came three days after the model became widely available to paid API users.
