A month ago, the Trump administration called Anthropic a national security risk and said it would no longer do business with the company. Last Friday, CEO Dario Amodei walked into the West Wing for a meeting with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent. The reversal wasn’t diplomatic—it’s technical. Mythos, an AI model Anthropic never intended to build for security work, had found thousands of vulnerabilities that no other tool could touch.
What Mythos Actually Does
Mythos Preview emerged from general improvements in reasoning and code generation—it wasn’t trained for offensive security. During internal testing, it discovered unpatched flaws across every major operating system and web browser. One OpenBSD bug had been sitting undetected for 27 years. An FFmpeg flaw went through 5 million automated test runs without triggering a single alert. Mythos found it in hours.
The total count: thousands of previously unknown, high-severity vulnerabilities.
This is the capability that changed the conversation in Washington.
The Glasswing Coalition
Anthropic didn’t release Mythos publicly. Instead, it assembled Project Glasswing—a closed group receiving access to the model in exchange for up to $100 million in cloud credits. The membership roster reads like a who’s-who of enterprise power: AWS, Apple, Cisco, Google, Microsoft, Nvidia, CrowdStrike, JPMorganChase. The stated purpose is offensive security—finding vulnerabilities before adversaries can exploit them.
Government agencies noticed. Intelligence services, CISA, and the Treasury Department all requested access. The Office of Management and Budget is preparing to grant federal agencies the ability to evaluate their own defenses using Mythos. National Cyber Director Sean Cairncross is leading a federal group to identify vulnerabilities in critical infrastructure using the model.
The Political Reversal
The juxtaposition is stark. Weeks before Amodei’s White House visit, the administration had declared Anthropic a “supply chain risk”—a designation typically reserved for foreign adversaries—and Trump himself said the government would not work with the company again.
A federal judge in San Francisco blocked enforcement of that directive. The Pentagon dispute remains active; Anthropic is barred from DoD contracts while litigation continues. But civilian agencies—Energy, Treasury—want Mythos capabilities and are unwilling to be collateral damage in the fight between the Pentagon and Anthropic.
Friday’s meeting was engineered specifically to separate those two threads. Progress with the White House and civilian agencies; no progress with the Department of War.
The Dual-Use Problem Nobody Can Solve
Here is the core tension that has no clean answer: the same capabilities that let Mythos find vulnerabilities also, theoretically, let adversaries find them first. Mythos could be used to breach US financial systems. It could also be used to harden them before bad actors strike.
This is no longer a technical problem. It’s a political one—and the political class is still figuring out how to think about it.
“It would be grossly irresponsible for the US government to deprive itself of the technological leaps that the new model presents. It would be a gift to China.”
That line, from a source close to the negotiations, captures the pressure facing both sides.
The Pentagon Problem
The DoD continues using Anthropic’s Claude models in active operations—including the war with Iran—despite the blacklisting. This contradiction hasn’t been resolved. The appeals court denied Anthropic’s request to temporarily block the Pentagon’s exclusion while a separate case proceeds. A San Francisco judge granted a preliminary injunction in a different case, but the DoD contract bar remains in place.
Anthropic hired lobbying firm Ballard Partners—where Susie Wiles worked for years—specifically to navigate the Department of War procurement issue. That relationship may prove useful; it hasn’t yet resolved the structural conflict.
What Comes Next
Three threads are moving simultaneously:
- Litigation continues through the courts, with the preliminary injunction providing some breathing room for Anthropic’s government business outside the DoD
- OMB access for federal agencies to test Mythos against their own systems is being structured—and could set precedent for how civilian agencies engage frontier AI models
- The Glasswing coalition is expanding as more agencies and companies request membership, increasing both the model’s reach and the scrutiny
The Washington door that seemed closed a month ago is open—but only partway, and only for now.
Related:
- Project Glasswing: Anthropic’s Restricted Access Model Program
- AI News — Daily coverage of the AI industry
