OpenAI restricts GPT 5.6 access to 20 pre-approved companies under White House security review

OpenAI restricts GPT 5.6 access to 20 pre-approved companies under White House security review

OpenAI is limiting access to its newest models following discussions with the Trump administration, citing national security concerns. The company says it's restricting GPT 5.6 to 20 pre-approved companies as a temporary measure while a White House review process unfolds, but cautioned that case-by-case government intervention shouldn't become standard practice.

In a blog post, OpenAI stated it hopes to make the model generally available within weeks and doesn't believe government approval processes should be the long-term default. "We don't believe this kind of government access process should become the long term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them," the company said.

The move follows a pattern. OpenAI previously restricted access to GPT 5.5 variants, including versions with cyber-attack capabilities, after White House discussions. Anthropic faced stricter enforcement: it halted all access to its Mythos model after the NSA flagged its ability to penetrate systems during a red team exercise. Anthropic has kept the model shuttered for two weeks.

The regulatory mechanism and its critics

The approach has drawn sharp criticism from AI safety advocates who call it the worst of both worlds. Rather than establishing clear, broadly applicable rules, the government is conducting high-touch, case-by-case reviews of specific model releases—giving regulators significant power over which companies can deploy which capabilities while offering no standardized compliance path.

The distillation problem

An immediate tension clouds these restrictions: the open-source frontier is advancing rapidly. Lab leaders say open-source models now lag frontier capability by approximately six months—a gap closing as researchers distill knowledge from closed systems into smaller, freely distributed models. If frontier models remain widely available through consumer subscriptions, large-scale distillation networks can extract advanced capabilities at scale, undermining the security rationale for restricted releases.

One alternative floated: as frontier models become more capable, they could detect distillation attacks directly rather than relying on access controls. Under this scenario, a ban list replaces a whitelist, with AI systems autonomously identifying and blocking suspicious query patterns. Most users would access widely available models; enforcement would happen in real time against identified threats. This approach would require constant iteration but sidestep the binary choice between open access and total restriction.

Bill Gurley has pushed this framing, arguing that if frontier models are intelligent enough to pose the security risks the government fears, they should be capable of defending against the extraction methods used to compromise them.

What remains unresolved

The restrictions implicitly assume a closed-source architecture can hold. If distillation continues unchecked, the policy framework collapses within months as capability parity arrives anyway. The administration has not publicly addressed how it plans to prevent mass-scale knowledge extraction or whether detection-based enforcement is viable at the model architecture level. OpenAI's stated timeline—weeks to general availability—suggests the company views the current restrictions as genuinely temporary, a bridge to clearer rules under Trump's AI oversight executive order.