Nearly 18,000 employees at the National Security Agency just discovered what foreign adversaries already knew: the White House is now the sole gatekeeper of frontier AI, and even America's most powerful intelligence agency does not get a free pass. According to two sources who spoke with CNBC and Politico, the NSA has lost direct access to Anthropic's Mythos 5, the company's most advanced cybersecurity model, and OpenAI was forced to hold one of its latest models for weeks before receiving government clearance to release it.
The developments represent the first concrete evidence that the White House's new frontier AI gatekeeping regime applies not just to foreign entities and private companies, but to the United States' own national security apparatus. If the NSA cannot get direct access to the most powerful AI models without White House approval, no one can.
The NSA Gap Opened When Project Glasswing Went Dark
Anthropic's Project Glasswing had been the primary channel through which the NSA and select government partners accessed Claude Mythos 5, the company's cybersecurity-specific model designed to detect zero-day exploits, analyze malware at machine speed, and generate defensive measures against nation-state threat actors. The NSA had been an active participant in Glasswing since its inception, using Mythos 5 to monitor critical infrastructure threats and adversarial cyber operations.
That access was revoked as part of the White House's broader clampdown on frontier model distribution. The administration, citing national security concerns, ordered a halt to new access grants across all frontier models until a formal review process could be established. The NSA, despite being a federal agency itself, was caught in the same net as everyone else. The absurdity is not lost on intelligence officials: the agency charged with protecting US networks from foreign cyber attacks cannot use the country's best cyber defense AI without permission from political appointees.
The situation has forced the NSA to rely on older, less capable models and in-house tools that lack the reasoning depth of Mythos 5. Intelligence sources say the degradation in threat detection capability is measurable, though they declined to provide specifics. Every day the review process drags on is a day that novel exploits and adversarial AI attacks go undetected.
OpenAI's Model Sat in Limbo for Weeks
The pipeline researcher's second source reveals that OpenAI found itself in a parallel predicament. One of the company's latest frontier models was held by the Trump administration for weeks before receiving clearance for release. The model, which sources describe as a significant leap in agentic capabilities, had completed internal safety evaluations and was ready for deployment when the White House intervened.
The hold lasted multiple weeks, not days, creating substantial operational and financial uncertainty for OpenAI. The company had already allocated compute resources, scheduled API access rollouts, and committed to enterprise customers who were waiting on the new capabilities. The delay forced OpenAI to push back product timelines and manage increasingly frustrated enterprise partners who had built their own roadmaps around the anticipated release.
The extended hold on OpenAI's model is particularly significant because it reveals that the White House review process is not a quick security check. It is a substantive, multi-week review that treats every frontier model release as a potential national security event. For startups and enterprises building on OpenAI's API, this means frontier model capabilities will arrive on government timelines, not product timelines. That is a structural risk that no amount of contractual SLA language can mitigate.
What This Means for Founders Building on Frontier AI
For founders and operators, the NSA's loss of access to Mythos 5 is the canary in the coal mine. If the United States government cannot guarantee its own intelligence agency access to frontier models, no startup should assume it has reliable access either. The era of assuming that frontier AI capabilities will be available on demand is over.
The practical implications are severe. First, any product or service that depends on a specific frontier model's capabilities now carries geopolitical risk. A White House review could interrupt your supply chain at any moment. Second, the cost of frontier model access is likely to rise as the government imposes compliance requirements on model providers, costs that will be passed through to customers. Third, the competitive landscape will bifurcate: companies that can navigate government access channels will have capabilities that others cannot reach.
The most important takeaway for founders is to diversify model dependencies now. If your entire product relies on OpenAI's latest frontier model or Anthropic's Mythos series, you have a single point of failure that is now controlled by political dynamics in Washington. Build abstraction layers, maintain fallback models, and consider whether your core functionality can be delivered with non-frontier models that operate outside the new gatekeeping regime.
The Precedent That Reshapes the Industry
This is the first time in American history that the executive branch has asserted direct control over access to a commercial technology product based on its capabilities rather than its compliance with existing laws. The closest historical parallel is the export control regime for nuclear technology, but even that was codified in legislation and operated through established agencies like the Department of Energy and the Nuclear Regulatory Commission. The current AI gatekeeping regime has no statutory basis and no established administrative process. It is being built in real time, which is precisely what makes it so unpredictable.
The NSA's loss of access to Mythos 5 will accelerate the already growing bipartisan concern in Congress about the administration's approach to AI governance. Multiple senators from both parties have privately expressed alarm that the White House is operating outside any legal framework. The coming legislative battles over the AI Regulatory Commission and the CREATE AI Act will define whether this executive-driven approach becomes permanent or gets codified into something more transparent and predictable.
For the rest of the industry, the message is unmistakable. Frontier AI is now a matter of national security policy, not product strategy. The companies that treat it as the former will survive the transition. The companies that treat it as the latter will not.
