The Trump administration is taking a more direct role in determining which outside organizations receive early access to powerful artificial intelligence models, potentially shifting a consequential part of the release process away from AI developers, according to reports from CNBC and Startup Fortune.
CNBC, citing two people familiar with the matter, reported that the White House has begun dictating which companies and other entities may work with some forthcoming frontier systems. Decisions about pre-release access have traditionally been made by model developers such as OpenAI and Anthropic, which have selected enterprise customers, government agencies and security researchers to test advanced systems before broader deployment.
The White House disputed the characterization that it has assumed approval authority over private-sector releases. An administration official told CNBC that companies participate voluntarily in government testing and consultations and retain control over when and how their models are introduced.
“Decisions on timing and scope of releases rest entirely with the companies,” the official told CNBC.
The disagreement reflects a distinction between formal legal control and the influence the federal government can exert through national-security reviews, export restrictions and requests for cooperation. Although the administration describes its model-testing framework as voluntary, CNBC’s sources said companies will need explicit government approval for the partners involved in certain controlled rollouts. Startup Fortune similarly reported that federal officials are becoming involved in selecting which outside parties receive access before public releases.
The emerging framework follows an executive order signed by President Donald Trump in June. According to Startup Fortune, the order asks developers to provide federal agencies with as much as 30 days of advance access to systems designated as “covered frontier models.” The order also gives agencies a role in evaluating potential outside partners, the outlet reported.
Frontier models are generally understood to be the most capable general-purpose AI systems available at a given time. Early access can be commercially valuable because it lets software companies prepare integrations, allows security specialists to test safeguards and helps government agencies assess potential threats. Control over that access can therefore influence which businesses are ready to use a model at launch and which researchers can examine it before deployment.
Cybersecurity appears to be a central concern. Advanced models may help defenders inspect code, identify vulnerabilities and automate parts of incident response. The same capabilities can also be misused to generate malicious code, accelerate vulnerability discovery or assist less-skilled attackers. Those dual-use risks have encouraged governments and AI laboratories to develop staged-release programs rather than making every advanced capability immediately and universally available.
Anthropic has used a partner initiative called Project Glasswing to provide limited access to its Mythos cybersecurity model, while OpenAI operates a comparable consortium called Daybreak, according to CNBC. The administration asked OpenAI to restrict access to its GPT-5.6 release, the outlet reported, and OpenAI subsequently said it would make new models available only to trusted partners in response to government requests.
The government’s growing involvement could change the future of those company-led programs. CNBC reported that the administration has launched a public-private cybersecurity initiative called Gold Eagle, envisioned as a clearinghouse for identifying and addressing vulnerabilities. One person familiar with the program told the outlet that it would place the White House at the center of decisions about which organizations may access new models.
Recent friction with Anthropic illustrates how government intervention can affect a model rollout. CNBC reported that the administration temporarily blocked access to Anthropic’s Claude Mythos 5 and Fable 5 models over national-security concerns before restoring it following weeks of negotiations. Startup Fortune provided a more detailed account, reporting that the Commerce Department imposed restrictions after Amazon researchers identified a way to prompt one of the models to produce exploit code for an actual software vulnerability.
According to Startup Fortune, Anthropic challenged both the government’s assessment of the problem and the transparency of the review process. The outlet reported that the dispute lasted 19 days, with the Commerce Department lifting its controls on June 30 and Anthropic beginning to restore broader access on July 1. CNBC described the restrictions and their eventual reversal but did not include all of those details.
The policy may not affect every advanced model in the same way. Startup Fortune reported that Google released Gemini 3.5 Pro without comparable restrictions because it did not meet the government’s capability threshold for a covered model. If that approach continues, federal scrutiny would be applied system by system, based on assessed capabilities rather than through a uniform rule covering every major developer or release.
That model-specific process could create uncertainty for businesses that rely on timely access to new APIs. Developers may have to account not only for a laboratory’s testing schedule but also for a federal review period and the possibility of a longer dispute. Smaller AI companies and application developers could be particularly sensitive to delays if their product launches depend on capabilities promised in a forthcoming model.
The administration is also confronting a broader strategic trade-off. Tighter controls may reduce the chance that powerful cyber capabilities are misused or transferred to risky actors. At the same time, restrictions on US companies could slow domestic deployment while overseas developers continue releasing less restricted systems, including open-weight models that users can run or adapt independently.
CNBC pointed to Chinese developer Moonshot AI’s Kimi K3 as evidence that the international capability gap is narrowing. The outlet reported that the model approached the performance of leading US systems and exceeded them on at least one independent benchmark. Such comparisons can vary substantially by test, but the arrival of increasingly capable Chinese models complicates efforts to control advanced AI solely through restrictions on American laboratories.
David Sacks, identified by CNBC as the former White House AI czar, warned that burdensome domestic controls could weaken the US position. The administration, however, said its objective is to work with frontier laboratories on security without suppressing innovation.
The central governance question is now whether the White House’s role remains an advisory and testing function or develops into an effective gatekeeping system. The administration maintains that private companies retain final release authority, while CNBC’s sources describe a process in which government consent is becoming necessary for certain early-access partnerships. How that tension is resolved will affect not only OpenAI, Anthropic and Google, but also the customers, researchers and security teams seeking access to their most capable models.
Sources: AI executive order, Anthropic