In February, the Pentagon decided Anthropic was too dangerous to trust. In April, the Treasury decided Anthropic was too dangerous to avoid.

Six weeks.

The February story is already documented. Defense Secretary Pete Hegseth gave Dario Amodei a Friday deadline. Drop the ban on fully autonomous weapons and the ban on mass surveillance of US citizens, or lose a $200 million defense contract. Amodei refused. Within hours the company was designated a "supply chain risk to national security," a phrase normally reserved for hostile foreign actors. Trump ordered federal agencies to stop using Anthropic technology, with a six-month phase-out window for the Pentagon itself. OpenAI signed the deal Anthropic wouldn't.

That was the administration's public position on the company. It still is.

On April 10, Scott Bessent and Jerome Powell summoned five bank CEOs to Treasury to discuss Claude Mythos, the Anthropic model that had launched three days earlier under the Project Glasswing programme. The recommendation was that banks consider using it for defensive vulnerability work. Four days later, Bloomberg reported that Treasury CIO Sam Corcos had gone further. He wasn't asking for a briefing. He was asking Anthropic for access to the model itself, so Treasury could run its own vulnerability tests. He hoped to have it, per the reporting, "as soon as this week."

Summoning CEOs is a warning. Asking for access is procurement reconnaissance. You don't request a working copy of a model unless you're thinking about using it, or thinking about understanding it well enough to regulate it. Either answer requires Treasury to be in active technical conversation with a vendor the administration has formally declared untrustworthy.

The easy reading is division of labor. Pentagon handles weapons and surveillance; Treasury handles financial stability; the agencies can disagree on the same company because they're optimising for different risks. From inside each building both calls look rational. Hegseth wanted Anthropic to remove safety features it considered load-bearing. Bessent and Powell want Anthropic to help defend the US financial system against a capability Anthropic itself warned about. No contradiction, just specialisation.

The harder reading is that "supply chain risk" means something. In February, the objection wasn't that Anthropic's technology didn't work. It was that the values embedded in the product — the specific guardrails Anthropic refused to remove — made the company unfit for government business. Those guardrails are still there. If they rendered the company unfit in February they render it unfit now. Treasury asking five banks to consider the technology, and then asking the vendor for a copy, doesn't unbrand the company. It ignores the brand.

There's a third reading worth naming, which the skeptics have been making for a week. Bruce Schneier called Glasswing a PR play. Alex Stamos called the Mythos framing "marketing schtick." AISLE replicated the headline findings with a 3.6-billion-parameter open-weight model costing eleven cents per million tokens. If they're right, then both the February blacklist and the April summoning are overreactions. One kind of overreaction got Anthropic banned from federal agencies. A different kind of overreaction is now getting its model briefed to the largest banks in the country, with access potentially approved for Treasury's own staff. The administration hasn't changed its mind about the company. It just changes which version of the company it's talking to.

Nothing has been retracted. The supply-chain designation stands. The phase-out order stands. The briefing happened. The access request is open. An AI policy reader trying to make the two positions cohere has to pick one, and the Trump administration has been remarkably unbothered about which one you pick.

Whichever you choose, the other one is still government policy.

Sources: