White House wants OpenAI to restrict GPT-5.6's initial release: Report

The US government is reportedly taking an unusually hands-on approach to OpenAI's next frontier AI model, signalling a shift in how advanced artificial intelligence is being deployed. Rather than allowing a broad rollout, GPT-5.6 could initially be made available only to a small group of approved customers, reflecting increasing concern over the security implications of cutting-edge AI systems.

According to The Information, OpenAI plans to launch GPT-5.6 through a restricted preview programme after the Trump administration asked the company to tightly control its initial release. The publication reported that the model will first be shared with a select group of close partners before a wider rollout is considered.

More from Tech

OpenAI's next model may launch with restricted access

During an internal meeting this week, OpenAI CEO Sam Altman reportedly told employees that the government would be "approving access customer by customer" during the preview phase, The Information reported. If the process proceeds as planned, OpenAI hopes to broaden availability "a couple of weeks later".

Neither OpenAI nor the White House has publicly commented on the reported arrangement.

If accurate, the move would represent a notable departure from OpenAI's previous launch strategy. While access to some earlier models has been staggered, the company has generally expanded availability relatively quickly through ChatGPT and its developer APIs. GPT-5.6, however, could become one of the first frontier AI models to undergo a government-supervised preview before reaching a wider audience.

The reported decision also highlights a broader change in Washington's approach to AI. Policymakers are increasingly viewing frontier models not just as commercial products but as technologies that could have implications for national security, cybersecurity and critical infrastructure.

Anthropic's Mythos and Fable put frontier AI under the spotlight

The reported restrictions on GPT-5.6 come shortly after The Information revealed details about two advanced AI projects being developed by Anthropic, codenamed Mythos and Fable.

According to the publication, the models are designed for cybersecurity tasks and are capable of identifying software vulnerabilities far more quickly than human researchers. They can also reportedly generate ways to exploit those weaknesses, raising concerns that such systems could be used for offensive cyber operations if deployed without sufficient safeguards.

The US government has reportedly restricted access to those Anthropic models as well, limiting them to a small number of trusted users because of their potential national security implications.

Cybersecurity experts have long argued that AI presents a double-edged sword. Models capable of discovering hidden flaws in software could help organisations patch vulnerabilities before hackers find them. At the same time, the very same capabilities could make it easier for malicious actors to compromise enterprise systems if access is not carefully managed.

Because Mythos and Fable remain unavailable to the public, their capabilities have not been independently verified, making it difficult to assess the extent of the risks they pose. Nevertheless, the reported restrictions on both Anthropic's cyber-focused models and OpenAI's GPT-5.6 suggest that governments are becoming increasingly cautious about how the most capable AI systems are introduced.

If the reported plans move forward, GPT-5.6 may mark the beginning of a new era in frontier AI, one in which the most advanced models are released gradually under closer government oversight rather than through the broad public launches that have defined the generative AI boom over the past few years.