White House pushes back on Anthropic's Mythos expansion

The White House is caught in an awkward staring contest with Anthropic, and it all comes down to a terrifyingly powerful model named Mythos.

According to reports from The Wall Street Journal and Bloomberg, Anthropic planned to more than double the number of private entities testing Mythos, expanding access from roughly 50 firms to about 120. But administration officials stepped in and said no.

Too scarce for comfort, too dangerous for the crowd

Officials cited two main headaches: compute bottlenecks and severe cybersecurity risks.

First, Washington is worried Anthropic simply does not have enough compute power to serve dozens of commercial clients without degrading performance for Uncle Sam. In short, the government wants dibs on the servers.

Second, Mythos is a different beast entirely. Introduced in early April, the model can automatically discover and exploit vulnerabilities in critical enterprise software. Anthropic itself deemed Mythos far too dangerous for a broad public rollout, opting instead for a walled-garden enterprise pilot.

Security concerns are not theoretical, either. Bloomberg previously reported that a small group of unauthorized users inside a private online forum managed to breach Mythos on the very day Anthropic announced its limited release.

Bashing the leadership while hoarding the weights

The standoff highlights a bizarre split inside the Trump administration.

On one side, Defense Secretary Pete Hegseth recently lashed out at Anthropic on Thursday, claiming the company is led by an ideological lunatic. The Pentagon spent months escalating a feud with the startup, pushing restrictions that triggered an active legal brawl.

On the other side, Axios reported that an upcoming White House AI memo will push agencies toward a multi-vendor strategy, essentially carving out a workaround so federal teams can bypass supply chain designations and use Anthropic anyway.

The White House wants to keep Mythos close for national security, block the private sector from hogging precious GPUs, and still avoid looking like it backed down to a company it publicly vilified.

Meanwhile, the window for holding this monopoly is closing fast. OpenAI's GPT 5.5 has reportedly matched Mythos in offensive cyber capabilities, and former AI czar David Sacks noted that all frontier models will likely hit this threshold within six months.

Why it matters

For enterprise security teams waiting in line, this is a clear warning sign. If you were banking on Mythos to automate defensive patching or red-teaming this quarter, do not hold your breath. Frontier-grade offensive tooling is turning into controlled technology, and commercial access will remain throttled.

For the broader industry, the government's justification sets a murky precedent. Blocking a private vendor from onboarding commercial customers over federal compute resource priority looks uncomfortably close to peacetime technological conscription.

In our view, Washington's strategy looks increasingly untenable. Throttling Anthropic might buy bureaucrats a few weeks of server priority, but with GPT 5.5 already at parity and other frontier labs just months behind, gatekeeping one vendor will not stop offensive AI from arriving everywhere else.