The White House rolls out the “Golden Eagle Program” to catch software vulnerabilities. CNBC reveals an inside story: OpenAI and Anthropic’s new models may require government approval.

The White House on Monday (7/14) launched a codename “Gold Eagle” network security initiative to coordinate between the government and the public, using cutting-edge AI models to find and patch software vulnerabilities before hackers do. But, according to a CNBC report on Friday (7/17), the real catch of the program is internal: it could give the White House the power to decide which companies can access the latest AI models early. OpenAI and Anthropic, when selecting their first batch of partners going forward, may have to obtain explicit government approval first. The White House denied to CNBC that it has any approval authority, stressing that corporate participation in testing is voluntary.
(Background: Part of the Trump administration’s easing of Mythos 5; OpenAI GPT-5.6 Sol is limited to White House-approved customers)
(Additional context: Trump signed an executive order stating AI companies can “voluntarily” submit the latest models to government review before release—did Mythos scare the White House?)

Table of contents

Toggle

• The obvious line is an AI漏洞-hunting machine
• The hidden line is what CNBC sees as the real focus
• The White House is rushing to distance itself from the words “approval”

Key takeaways

• The White House on July 14 rolled out “Gold Eagle,” using cutting-edge AI models to coordinate the identification and patching of open-source software vulnerabilities.
• CNBC reveals that this program could let the White House control which companies are allowed to access cutting-edge AI models early.
• Since June 26, GPT-5.6 has been open only to about 20 partner organizations that have undergone government review; the White House denies it has approval authority.

On Monday (7/14), the White House announced the “Gold Eagle” program jointly with Treasury Secretary Scott Bessent and Department of Homeland Security Secretary Markwayne Mullin. This is an AI-driven vulnerability coordination hub. The source traces back to an executive order signed by Trump in early June, aiming to identify, fix, and push patches to users before the vulnerabilities unearthed by AI are exploited by bad actors.

The reason for this program is that the number of vulnerabilities AI can uncover in a day is so large that the cybersecurity community simply can’t patch them all. Rather than having everyone rush to respond on their own, the government decided to build a central switchboard to schedule work—this is the rationale for Gold Eagle presented on the record.

The obvious line is the AI vulnerability-hunting machine

The technical core of Gold Eagle is the “VINCE” system (Vulnerability Information and Coordination Environment), developed in cooperation between the White House and the Software Engineering Institute (SEI) at Carnegie Mellon University. Its job is to receive vulnerability reports found by third parties using AI, then prioritize which ones should be patched first.

The government says the system has already started operating and collecting intelligence. The lead is the Treasury Department, with CISA, the Department of Homeland Security, and the Department of Defense all involved. The focus is on open-source software, because these codebases underpin a large amount of critical infrastructure but are often never reviewed carefully. And the model used to hunt for vulnerabilities is a cutting-edge one like Anthropic’s Mythos.

The hidden line is the real focus in CNBC’s eyes

But citing an additional angle from people familiar with the matter, CNBC suggests something even more intriguing than vulnerability hunting. This coordination hub may also be responsible for vetting which companies and institutions can access cutting-edge AI models early—effectively keeping the “ticket” for releasing models in the hands of the government. For future partner selection, OpenAI and Anthropic may both have to secure explicit government approval first.

This is not pure speculation. The U.S. government has already intervened in screening the first batch of GPT-5.6 partners. OpenAI began phased releases starting June 26: at first, it opened access only to about 20 partner organizations that had been reviewed by the government, rather than launching directly to the public at full scale. OpenAI also had a cybersecurity model alliance called Daybreak in its portfolio.

Anthropic has also had a rough time. Its Claude Mythos 5 and Fable 5 were at one point blocked by the Trump administration over “national security concerns,” with access restored only after negotiations lasting weeks.

The White House is quick to distance itself from the words “approval”

In response to claims about power transfer, the White House spoke firmly to CNBC. Officials denied having any approval authority, emphasized that it did not give OpenAI a “green light,” approve, or clear anything, that it did not need such permissions, and said the timing and scope of model releases are “entirely decided by the companies.”**

The White House also reiterated that corporate participation in testing remains voluntary, and that when to release and to whom to open it are ultimately decided by the companies themselves.

The White House’s claim has support. The executive order dated June 2 from Trump proposed a voluntary security testing framework, giving federal agencies a 30-day access window before a model is publicly released. The order explicitly states that it does not include “mandatory” government licensing, pre-review, or permission requirements.

Frequently asked questions

What is the White House’s “Gold Eagle” program?

Gold Eagle is an AI cybersecurity coordination center the White House plans to launch in July 2026. It is led by the Treasury Department and, through the VINCE system co-developed with Carnegie Mellon University, coordinates between government and the public using cutting-edge AI models to find and patch vulnerabilities in open-source software.

Can the White House really decide when OpenAI and Anthropic release AI models?

The White House denies it has approval authority and stresses that corporate participation in testing is voluntary, while the timing and targets of release are determined by the companies. But in practice, GPT-5.6 has already been limited to about 20 government-reviewed partner organizations, and Anthropic’s models have also been blocked on national security grounds.