The official version looks clean. Anthropic developed a model so powerful that institutions had to step in. The Pentagon flagged it as a supply chain risk. The White House pulled it from federal systems. Trump called it woke and demanded its removal. The story has a corporate villain and a government protecting national security. The following ninety days completely dismantle that narrative.

The Pentagon's initial stance had concrete reasons behind it. Anthropic refused to let its models be used for all legal purposes, including autonomous weapons and mass domestic surveillance. From the military's perspective, that restriction is unacceptable. The chain of command cannot depend on a private company's ethical values. It's a real governance problem, not institutional paranoia.

The data on Mythos Preview is verifiable. The UK's AI Safety Institute confirmed that the model completes thirty-two-step corporate attack simulations—tasks that take human experts twenty hours—in three out of ten attempts. On expert-level tasks it succeeds seventy-three percent of the time. That capability shifts the strategic calculus for any agency protecting critical infrastructure.

The same day the Pentagon's CTO declared that Anthropic remained a risk, the department announced agreements with seven companies to deploy artificial intelligence on classified networks: SpaceX, OpenAI, Google, Microsoft, Nvidia, Amazon Web Services, and Reflection. Anthropic doesn't appear on the list. Meanwhile, the NSA is using Mythos Preview to scan for vulnerabilities. The right hand litigates. The left hand deploys.

A source inside the Trump administration summed it up with a bluntness no official statement would ever admit: every intelligence agency uses Anthropic. Every agency except War wants it because Anthropic avoids contributing to killing people, and War's position is don't tell us what to do. But if you work at the Department of Energy, the Chinese attacking the power grid matters more than those reservations. So you go looking for Anthropic.

What rarely shows up in the coverage is the question of exclusivity. The NCC Group experts who evaluated the model were direct about it: this capability will be replicated by other artificial intelligences within months. The Council on Foreign Relations noted that advanced model code leaks frequently and that leading companies show a clear tendency to replicate rival capabilities, typically within months. Peter Swire, a former adviser to Clinton and Obama, described it as, above all, a public relations success.

The capability is real. The qualitative leap from what came before is more modest than the press releases suggest. This matters because it transforms the negotiating logic. The White House is now convening companies across several sectors to prepare a possible executive action reversing the OMB directive and reinstating Anthropic. Internal sources describe the effort as saving face and bringing them back. The company Trump called woke is negotiating its return from a position of strength it didn't have sixty days ago. The red line on autonomous weapons remains unresolved. What changed is who needs whom more.

I've seen this dynamic play out in different contexts. When a tool becomes indispensable, the institutions trying to regulate it lose real enforcement power. Incentives realign and threats lose credibility. The Pentagon arrived at that point with Mythos. Not out of conviction, but out of technical dependence. How far this reversal of power will go is something I still can't say with certainty.

There's a detail almost absent from the headlines that turns out to be revealing. The club of forty organizations with access to Mythos includes members Anthropic never publicly named, among them the NSA and the UK's AI Safety Institute. At least one Discord group got in uninvited on launch day using contractor credentials and URL guessing. The company officially declared a national security threat doesn't know exactly who is using its most advanced model. The question of who audits Anthropic remains open while they negotiate the terms of reinstatement.

In February, the Pentagon equated Anthropic with risks associated with foreign adversaries. In May, the White House is looking for a way to bring them back without losing face. The model hasn't changed. The company's ethical principles haven't changed either. What changed is the scale of the dependence.

The fact that closes the loop is this: the company formally excluded from the Pentagon's classified artificial intelligence agreements is the one every intelligence agency wants to use. The one that was included—OpenAI—already holds military contracts without the ethical restrictions that made Anthropic famous. And the capability that turned Mythos into a strategic asset will be replicated by competitors within months. The government traded regulatory legitimacy for that temporary exclusivity. This is exactly what happens when technology outpaces the institutions supposedly regulating it.

Certain historical powers ended up depending on innovations they publicly rejected. The parallels suggest that technical dependencies rewrite the rules of authority faster than any official statement will admit. How this will play out in the coming years is the question that won't leave me alone.

What does this reversal of power reveal about who really controls advanced technology?

Sources:

1. Al Jazeera — Pentagon announcement of agreements with seven AI companies for classified networks

2. Time — Statement from an anonymous source within the Trump administration on intelligence agencies' use of Anthropic

3. NBC News — Results from the UK's AI Safety Institute on Mythos Preview's capabilities

4. MSN / The Sun — Statements from NCC Group and the Council on Foreign Relations on capability replication

5. ScienceInsights — White House efforts to reverse the OMB directive on Anthropic