President Donald Trump told CNBC on Tuesday that a Pentagon deal with Anthropic is "possible," a dramatic reversal from weeks ago when his administration blacklisted the AI company as a national security risk. The shift comes days after Anthropic CEO Dario Amodei met with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent in what the White House described as a "productive and constructive" meeting.

The topic of that meeting: Mythos, Anthropic's most powerful AI model and what cybersecurity experts are calling a "watershed" capability.

How We Got Here

The relationship between the Trump administration and Anthropic cratered earlier this year when Anthropic refused to remove AI safeguards for use in autonomous weapons systems and surveillance. The Pentagon, under Secretary Pete Hegseth, demanded "unfettered access" to Claude "for all lawful purposes," including during wartime operations. Anthropic said no.

The fallout was swift. The administration blacklisted Anthropic. Trump issued a social media directive against the company. Anthropic sued the Trump administration in both San Francisco and Washington, D.C., to reverse the blacklisting. A federal judge temporarily blocked Trump's social media directive. Meanwhile, the Pentagon continued using Claude anyway, particularly during ongoing operations in Iran.

Enter Mythos

Then Anthropic announced Mythos. The model's capabilities, while not fully public, have been described by Reuters as raising significant concerns. The White House held discussions with Amodei about "collaboration, cybersecurity, and balancing AI innovation with safety." The NSA already has access to Mythos. CISA, the agency that protects American infrastructure from cyberattacks, is locked out.

The dynamic is clear: the administration needs Anthropic more than it wants to admit. Mythos represents capabilities that the military and intelligence community want access to. The question was always whether Anthropic would bend its safety principles to get back in the government's good graces.

What 'Shaping Up' Actually Means

Trump's language is revealing. "Shaping up" implies Anthropic is moving toward the administration's position, not the other way around. Whether Anthropic has actually agreed to loosen safeguards, offered some compromise on military access, or simply showed up willing to talk is not yet public. But the framing tells you who Trump thinks is winning this negotiation.

This matters beyond the Pentagon deal. Anthropic is widely expected to IPO in the near term. Its lobbying spend on Washington has already surpassed OpenAI's. Being blacklisted by the U.S. government is not the kind of risk factor you want in an S-1. A thaw with the White House is existentially important for the company's public market ambitions.

The Safety Question

Anthropic built its brand on being the "safety-first" AI lab. Dario Amodei has spent years arguing that AI capabilities must come with robust guardrails. The company's Responsible Scaling Policy was supposed to be the industry gold standard. Now the company is in active negotiations to give the Pentagon access to its most powerful model.

There is a version of this that ends well: Anthropic negotiates meaningful use restrictions, the military gets access with guardrails, and the company maintains its safety credibility. There is also a version where "shaping up" means capitulation. We do not know which version we are in yet. But the fact that Trump is smiling should make the safety community nervous.