Anthropic and the White House: a nine-day thaw

Nine days

On April 8, a federal appeals court in Washington declined to suspend the Pentagon's designation of Anthropic as a "supply-chain risk." On April 17, Dario Amodei, CEO of Anthropic, walked into the White House for a meeting the Trump administration would later describe as "productive and constructive."

Nine days between the two.

When we covered the Anthropic blacklist ten days ago, the most credible scenario was a long legal battle pitting an isolated company against an irritated administration. Nine days later, the scenario has changed. No ideological reversal, no official olive branch. Just an administration realizing it needs the very technology it just banned.

What happened between the two dates

Things accelerated around April 7. Scott Bessent, Treasury Secretary, and Jerome Powell, chair of the Federal Reserve, summoned the CEOs of the largest US banks to an urgent meeting. The topic: the cybersecurity risks posed by Mythos, Anthropic's new model.

Mythos is not a consumer chatbot. It is a model capable of identifying zero-day vulnerabilities in software. In plain terms: it finds flaws that humans and conventional tools have never caught. Anthropic deployed it in limited capacity through Project Glasswing, with JPMorgan, Apple, Google, Microsoft and Nvidia as initial partners.

The banks are testing the model in-house. Goldman Sachs, Citigroup, Bank of America, Morgan Stanley are all getting access over the following days. The same Trump administration that officially blacklisted Anthropic at the Pentagon is encouraging Wall Street to use Anthropic's technology to protect itself.

Two doors in the same house. One slammed shut on the Pentagon side. One held open on the Treasury side.

"Every agency except DOD"

On April 17, Amodei met Susie Wiles, chief of staff at the White House, and Scott Bessent. Not at the Pentagon, not shaking hands with Pete Hegseth. The meeting took place elsewhere, with different players.

A senior administration source quoted by Axios sums it up: "every agency except DOD wants to use the company's technology." Every federal agency except the Department of Defense wants to use Anthropic's tools.

The internal split is stark. And the signal to the AI ecosystem is clear: a Pentagon blacklist is not a blanket ban from the US government. It is a sector-specific conflict.

One detail that matters: when Trump was asked what he thought of the Amodei meeting, he answered "Who?" and claimed he had not been briefed. The thaw is being driven by the second tier, not the president. That is information in itself about where power actually sits.

Anthropic's strategy: work both tables at once

On April 13, Jack Clark, Anthropic co-founder and Head of Public Benefit, spoke at the Semafor World Economy summit. He called the Pentagon dispute a "narrow contracting dispute." A minor contractual disagreement. Nothing more.

In the same appearance, he added: "Our position is that the government has to know about this stuff, and we have to find new ways for the government to partner with a private sector that builds tools that are transforming the economy but have national security implications."

The line is precisely calibrated. Downplay the legal fight (nothing serious), reaffirm openness to dialogue (we are responsible partners), keep the lawsuit alive (we are not conceding on substance).

Anthropic is playing multi-level chess. The Pentagon blacklist is still in place. The lawsuit is still moving forward. The core disagreement (no fully autonomous weapons, no mass surveillance of US citizens) remains untouched. And meanwhile, the CEO is discussing cybersecurity and the AI race at the White House.

What the thaw does not resolve

One is entitled to ask what actually changed. On ideology: nothing. The administration initially wanted Claude "without restrictions" for military use. Anthropic refused. That disagreement was not resolved at the White House on April 17.

What changed is Anthropic's perceived usefulness to the administration. Mythos arrived, and its ability to flag critical vulnerabilities shifted the equation. You do not entirely blacklist a company you need to secure the country's systemic banks.

The tacit compromise is probably this one: the Pentagon keeps its hard line (Anthropic remains blacklisted for military use), but other agencies and Treasury get normal access. Anthropic does not concede on its terms, but accepts official dialogue. Everyone saves face.

The lesson for the AI ecosystem

For other labs watching, the signal sent is instructive. If you have technology the government cannot afford to ignore, you can hold a firm ethical line and survive a sector-specific ban. If your technology is replaceable, you do not have that leverage.

OpenAI, Google and Meta have not publicly adopted the same conditions as Anthropic (explicit refusal of autonomous weapons, refusal of mass surveillance). Why not? Not necessarily out of opposite conviction. Possibly because they cannot afford a political clash like this one. Or because they have not yet had to choose.

Anthropic chose, took the hit, and is starting to see the backlash ease. It is a position that, if it holds over the long term, could redefine what is expected of an AI lab supplying governments. Contract equals conditions. And some conditions are not negotiable.

What we are watching next

The Anthropic vs. DOD case on the merits is still pending. The constitutional arguments on the First and Fifth Amendments have not yet been ruled on. The final judgment could lift the blacklist or confirm it.

Meanwhile, we are watching whether the Pentagon joins the movement. If Mythos becomes indispensable for national cybersecurity, Hegseth's hard line becomes politically harder to hold. The thaw, if confirmed, will not be announced in a press release. It will show up in the contracts signed or not signed over the coming weeks.

To follow this seriously, keep an eye on three indicators: new Anthropic federal contracts (outside DOD), court filings in the lawsuit, and any public statement from Hegseth or Trump on the topic. Everything else is noise.