Loading page…
The proposed arrangement puts national-security testing nearer to frontier-model releases without creating a licensing system. Whether leading developers will participate—and on what terms—remains unsettled.
Listen to this story
A proposed federal testing channel would let AI developers hand qualifying models to the government for up to 30 days before sharing them with other trusted partners. The framework follows a June executive order requiring classified cyber-capability tests, but it is explicitly not a pre-release approval regime. NSA officials are seeking broad commercial access and will help determine which models qualify.
Participation remains voluntary; the June order bars mandatory licensing or approval for new AI models.
NSA Director Tim Kosiba will help decide which systems qualify, alongside ONCD, CISA and other agencies.
The benchmark is classified, limiting public visibility into attendees, qualifying models and test results.
The White House has invited OpenAI, Anthropic, Meta and Google to review a voluntary framework for cybersecurity testing of frontier AI models. At the same time, the National Security Agency is seeking access across the commercial AI market, placing the intelligence agency closer to the models the framework is meant to scrutinize before release.
The policy effort began with a June executive order directing national-security agencies to develop classified tests for identifying AI systems with hacking capabilities serious enough to need added scrutiny. It sets up a voluntary arrangement: developers may provide qualifying models to the federal government for as long as 30 days before releasing them to other trusted partners.
That structure is deliberately not a federal approval gate. The order prohibits a mandatory licensing or approval process for new AI models, leaving participation to developers. The White House meeting is intended to let the four companies review the finished framework, its classified benchmark and the next implementation steps.
NSA Deputy Director Tim Kosiba said the agency wants access to all commercial AI models and is in extensive conversations with frontier-model companies. A related June national-security directive also told intelligence and defense agencies to build proactive industry partnerships and draw advanced models from a diverse supplier base.
The NSA director has a formal role in the new program: deciding which systems count as covered frontier models, in consultation with the Office of the National Cyber Director, the Cybersecurity and Infrastructure Security Agency and other government offices. That decision determines which models can enter the voluntary pre-release channel.
Kosiba did not name the companies in those discussions, identify the models the NSA already uses or confirm that any developer has granted access to an unreleased system. The agency’s public comments therefore do not establish that the pre-release program is operating or that companies have begun supplying models through it.
Government access can still arrive after a public launch. OpenAI’s GPT-5.6 models became available to federal customers through its FedRAMP-authorized enterprise service this month, after their July public release. The new framework is aimed at an earlier point: testing qualifying systems before they reach other trusted partners.
The benchmark’s classified status creates a built-in visibility limit. Outsiders may not know who attended the framework review, and the public record has not identified participating developers or a qualifying unreleased model. The next consequential move is whether the voluntary design produces a disclosed handoff from a lab to the government.
Editorial analysis
The policy’s central test is not whether Washington can define a cyber benchmark in secret. It is whether leading labs see enough value in voluntary participation to bring sensitive systems into a government review process before release. The executive order rejects formal licensing, so cooperation is the mechanism rather than a fallback. Watch for a concrete sign of uptake: confirmation that a developer has supplied a qualifying unreleased model, or public detail on how the NSA director will apply the frontier-model threshold.
Loading discussion...
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.