The invoice landed in March: $130,000 for the services of Brian Ballard. In the grand scheme of Washington lobbying, it is a rounding error, but the client was Anthropic, and the target was the inner circle of the Trump administration. For a company that the Pentagon had recently branded a “supply chain risk,” the investment was more than just a line item; it was a bid for survival. Within weeks, the secret meetings began. On April 18, Anthropic CEO Dario Amodei was summoned to the White House not to discuss a ban, but to negotiate a deployment. The product in question, a cutting-edge model known as Claude Mythos, has become too dangerous to release to the public and too powerful for the federal government to ignore.
The situation reveals a deepening schism in American technology policy. On one side, the Department of Defense maintains a formal exclusion of Anthropic’s software from its workflows, citing concerns over the boundaries of use and the company's refusal to greenlight the development of fully autonomous weapons. On the other, the White House is currently bypassing its own hawkish rhetoric to integrate a “modified version” of Mythos into the Department of Energy, the Treasury, and Homeland Security. It is a classic Washington paradox: a technology deemed a risk to the military has become the primary shield for the civilian state.
The Glasswing Paradox
To understand why the White House is willing to go back on its word, one must look at Project Glasswing. When Anthropic launched this initiative in early April, it wasn’t pitching a better chatbot for writing emails. It was unveiling a system capable of identifying thousands of zero-day vulnerabilities in critical infrastructure code. In internal testing, Mythos demonstrated an uncanny ability to navigate complex software repositories and find the kind of architectural flaws that keep national security officials awake at night. For the first time, the speed of vulnerability discovery has outpaced the human capacity to patch them.
This is what engineers call a dual-use crisis. The same logic that allows an AI to identify a flaw in a power grid’s control software also provides the roadmap for a catastrophic attack. Anthropic has kept Mythos behind a “Gated Research Preview,” limiting access to a handful of partners like Amazon AWS, Microsoft, and Palo Alto Networks. But the White House realized that if these capabilities exist in the private sector, the state cannot afford to be the last to wield them. The demand for Mythos within federal agencies is not coming from office managers; it is coming from the people responsible for the physical integrity of the electrical grid and the financial system.
A Systemic Risk to the Ledger
The anxiety surrounding Mythos reached a fever pitch in early April when Treasury Secretary Scott Bessent and Federal Reserve Chairman Jerome Powell summoned the CEOs of Wall Street’s largest banks to Washington. This was not a routine briefing. The discussion centered on the potential for Mythos, or a competitor’s equivalent, to trigger systemic financial disruptions. The Securities Industry and Financial Markets Association (SIFMA) warned in an open letter that malicious use of such models could lead to large-scale identity theft or, more critically, the exploitation of high-frequency trading vulnerabilities that could crash markets in milliseconds.
From a technical perspective, the risk is not that the AI will “decide” to attack a bank. The risk is the erosion of the time-buffer that currently protects modern systems. Traditionally, cybersecurity is a game of cat and mouse where the defender has a slight home-field advantage. Mythos shifts that dynamic by automating the “search” phase of an attack. When a model can scan millions of lines of code and identify an exploitation chain in seconds, the defense line, which relies on a window of discovery and disclosure, effectively vanishes. This is why Powell and Bessent are treating the model not as a software tool, but as a macroeconomic variable.
The 'Modified' Straitjacket
The White House’s solution to the Pentagon ban is the creation of a “modified version” of Mythos. In the parlance of Brussels or Berlin, this would be seen as a desperate attempt at technological sovereignty through administrative tinkering. In Washington, it is a way to bypass procurement blacklists. This modification is two-fold. Technically, it involves hard-coded restrictions on the model's ability to output actionable exploit code, effectively turning it into a “read-only” security consultant. Institutionally, it locks the model within a closed federal circuit, managed by the Office of Management and Budget (OMB).
Federal Chief Information Officer Gregory Barbaccia has already begun the process of setting the boundaries. An internal memorandum indicates that agencies like the Department of Justice and the State Department will receive access, but only under a framework that requires exhaustive logging of every query. This is a far cry from the open-ended AI assistants marketed to the public. The government is essentially building a digital cage around the model, hoping to harness its diagnostic brilliance while neutralizing its offensive potential. Whether such a cage can actually hold a model with emergent capabilities remains a subject of intense debate among the very few researchers who have seen the full Mythos weights.
The European Lens: Sovereignty vs. Safety
For observers in the European Union, the Mythos saga is a cautionary tale about the reality of the AI Act versus the exigencies of real-world power. While the EU focuses on the classification of “high-risk” systems and transparency requirements, the United States is moving toward a model of state-captured AI development. By “modifying” and nationalizing the deployment of private-sector models, Washington is creating a precedent where the most powerful technologies bypass standard regulatory scrutiny under the umbrella of national security.
This creates a significant headache for German and French industrial policy. If the U.S. government is integrating these capabilities into its Treasury and Energy departments, European counterparts will find themselves at a structural disadvantage unless they can develop equivalent, sovereign models. The problem is that the European semiconductor and AI landscape remains fragmented. While firms like Mistral in France or Aleph Alpha in Germany aim for transparency and safety, they are competing with American entities that have essentially become an extension of the state security apparatus. The “modified version” of Mythos is a signal that the era of AI as a general-purpose SaaS product is ending for critical sectors. It is becoming a controlled substance.
Procurement as a Weapon
The friction between the Pentagon and the White House also highlights a failure in how the military-industrial complex handles modern software. The Pentagon’s refusal to use Anthropic stems from a desire for total control—specifically, the right to use models for autonomous weaponry. Anthropic’s refusal to comply is often framed as an ethical stance, but it is also a pragmatic business decision: being labeled a “death tech” firm would alienate the commercial partners like Amazon and Google that provide the massive compute necessary to train models like Mythos.
As the OMB prepares to roll out access to Mythos in the “coming weeks,” the focus will shift from the drama of the ban to the reality of the deployment. The government is gambling that it can domesticate a technology that was designed to be disruptive. History suggests that the bureaucracy is rarely faster than the code it tries to regulate. The White House has decided that the risk of using Mythos is high, but the risk of not using it is higher. It is the kind of progress that doesn’t fit on a slide deck, and it is a reality that Brussels will eventually have to confront, likely after the first zero-day is discovered by an AI that the Pentagon technically doesn't own.
Washington has accepted the reality that these capabilities will inevitably permeate the global infrastructure. The competition has shifted from prevention to domestication. The U.S. government has the model; now it just has to figure out which department gets to keep the keys.
Comments
No comments yet. Be the first!