Welcome back to The AI Mandate.

On Tuesday's Scan, I looked at how the biggest defense consultants and contractors are transforming their own businesses with AI. GDIT shipping an AI platform that runs from a laptop to orbit. BAE Systems embedding Scale AI's tools into combat vehicles. Lockheed Martin spinning off an entire subsidiary to sell AI infrastructure to government. To me, the takeaway is clear: the companies serving federal agencies are moving faster on AI than the agencies themselves.

This week's Intelligence Report is about the company caught in the middle of all of it. Anthropic built the AI model the military depends on in an active war, released a cybersecurity model so powerful they had to nerf it’s release, then watched its CEO walk into the White House while the Pentagon still calls it a national security threat.

This week, the full story on Anthropic, the Pentagon, and a conflict that keeps getting stranger.

Let's get into it.

In the first Intelligence Report, I covered the Anthropic-Pentagon conflict as one thread in a broader story about AI's role in the Iran campaign. Since then, the situation has moved in a direction that would be difficult to script. Court rulings, a new AI model with serious national security implications, a White House meeting, and a growing contradiction at the center of the government's own AI strategy.

Here's where things stand.

For readers just joining, here’s the quick version. Anthropic signed a $200 million contract with the Pentagon in July 2025, which made Claude the first frontier AI model deployed on the DOD's classified networks. By September, negotiations over expanded access broke down as Anthropic drew two red lines: no mass surveillance of Americans, no fully autonomous weapons without human oversight. Whereas, the Pentagon wanted unrestricted access for "all lawful purposes."

The fallout was swift. President Trump directed agencies to stop using Anthropic's technology. Defense Secretary Pete Hegseth labeled Anthropic a supply chain risk, a designation previously reserved for foreign adversaries like Huawei and Kaspersky. Anthropic filed two lawsuits challenging the designation, one in San Francisco federal court and one in the D.C. Circuit Court of Appeals.

On March 26, Judge Rita Lin in San Francisco granted Anthropic a preliminary injunction. Her 43-page ruling found that the government's stated reasons for the designation were pretextual and that the real motive was retaliation for Anthropic's public stance on AI guardrails. She called the administration's actions "Orwellian" and wrote that the government "went further" than simply choosing a different AI vendor. The ruling barred 17 federal agencies from enforcing the ban while the case proceeds.

The D.C. Circuit court, however, had a very different ruling. On April 8, a three-judge panel denied Anthropic's request to temporarily block the supply chain risk label under the separate statute the Pentagon cited. The court acknowledged Anthropic would "likely suffer some degree of irreparable harm," but ruled the company's interests were "primarily financial in nature" and didn't meet the threshold for emergency relief. The panel set oral arguments for May 19.

As I covered last week, Claude remains embedded in Palantir's Maven Smart System and actively supporting targeting operations in the Iran campaign. The Pentagon is fighting Anthropic in court while relying on its technology in a war zone. This contradiction was already notable. Now, Mythos shines an even brighter light on it.

On April 7, Anthropic announced Claude Mythos Preview, a new frontier model with cybersecurity capabilities that the company described as too dangerous for public release. During testing, Mythos identified thousands of zero-day vulnerabilities across every major operating system and web browser. It can autonomously discover, exploit, and in some cases reverse-engineer software flaws that were previously unknown to developers. The UK's AI Security Institute confirmed these capabilities are real, calling Mythos the first AI model able to complete its full network attack simulation.

Anthropic restricted access to roughly 40 organizations through Project Glasswing, an initiative designed to use Mythos for defensive cybersecurity while sharing findings openly across the industry. The partner list includes Amazon, Apple, Google, Microsoft, CrowdStrike, and Palo Alto Networks.

Then, this past Sunday, Axios reported that the NSA is using Mythos Preview inside its classified networks. The same Defense Department that labels Anthropic a supply chain risk now has one of its own agencies running the company's most powerful model. A third source told Axios that Mythos is being used even more broadly within the defense and intelligence community. The White House OMB is reportedly preparing safeguards to give additional agencies access.

On April 17, Anthropic CEO Dario Amodei went to the White House and met with Chief of Staff Susie Wiles, Treasury Secretary Scott Bessent, and National Cyber Director Sean Cairncross. The meeting was seemingly about Mythos and how the government should engage with its capabilities. Both sides called it productive.

This was the first high-level meeting between Anthropic leadership and the White House since the Pentagon designated the company a supply chain risk. Though when reporters asked President Trump about it on a runway in Phoenix, he responded: "Who?" and said he had "no idea" Amodei had been there.

Whether that's credible is beside the point. What matters is the direction. Parts of the administration are moving toward engagement while the Pentagon holds the legal line. Sources told Axios that next steps from the meeting would focus on how departments other than the Pentagon can access Anthropic's tools. And Trump did say yesterday on CNBC that Anthropic is “shaping up” and there’s a “possible deal” being discussed with the Pentagon to use their models again.

While the government fights Anthropic in court, the market is fully bought into Anthropic and its capabilities. Anthropic closed a $30 billion Series G in February at a $380 billion valuation. Annualized revenue reportedly hit $30 billion by end of March, up from $9 billion at the end of 2025. And in mid-April, Bloomberg reported that VCs have been offering Anthropic a preemptive funding round that could value the company at $800 billion or more! Anthropic hasn't taken them up on it yet, but the demand exists.

The supply chain risk designation was supposed to cripple the company. Eight weeks later, investors are lining up and driving the valuation to unimaginable heights.

The Anthropic-Pentagon conflict has evolved from a contract dispute into the most consequential test case in American AI governance. The question at the center is no longer whether Claude is good technology. The military's own operational reliance on it, from Maven to Mythos to Amodei’s Washington visits, answers that. The question is whether an AI company can set boundaries on how its technology gets used and still be a viable government supplier.

The government's own actions are undermining its legal argument. You cannot credibly argue that a company threatens national security while the NSA runs its most sensitive model and the military depends on it in an active theater of war. Judge Lin's ruling already flagged this contradiction. The May 19 oral arguments in the D.C. Circuit will test whether the second court sees it the same way.

Congress has been absent. No legislation. No hearings on the supply chain risk process. No framework for resolving disputes between the government and its AI suppliers. Their silence comes across as a policy choice, leaving the most important AI governance questions in America to be settled by executive orders and court injunctions.

This story is far from over and I’ll be covering all the twists and turns as they unfold.

A quick scan of how governments around the world are building, deploying, and betting on AI.

🇬🇧🇪🇺 UK evaluates Mythos, EU gets locked out. The UK's AI Security Institute published a technical evaluation of Mythos Preview within a week of Anthropic's announcement, finding it was the first AI model to complete a full 32-step network attack simulation. The UK government followed up with an open letter to business leaders warning that frontier AI cyber capabilities are now doubling every four months.

Meanwhile, the European Commission was not among the 40 organizations granted access to Mythos. AI safety groups have written to Brussels arguing the EU AI Office lacks the technical staff to evaluate models like Mythos, and demanding it quadruple its safety unit. The gulf between the UK's hands-on evaluation capacity and the EU's regulatory apparatus is widening in real time.

🇪🇪🇺🇦 Estonia and Ukraine are building governments that run on AI. Estonia is moving toward citizens interacting with their government primarily through AI agents, multimodal systems that can issue licenses, process benefits, and speak any language. Ukraine is pursuing the same playbook under wartime pressure. Both countries likely see AI as existential infrastructure for small nations facing a hostile neighbor with a population advantage.

🇿🇦 South Africa publishes its draft national AI policy. South Africa posted its Draft National AI Policy on April 10, opening a 60-day public comment window. The approach is worth watching: instead of a single AI authority, South Africa chose a sector-specific, multi-regulator model. A different bet than the EU's centralized framework, and one that may prove more adaptable.

🇸🇦🇦🇪 The Gulf keeps building. Saudi Arabia increased its AI investment pledge from $600 billion to $1 trillion during Crown Prince MBS's Washington visit. The UAE's Stargate project, a 1GW AI compute cluster backed by G42, OpenAI, Oracle, and NVIDIA, is the most ambitious single AI facility in the region. MENA tech spending is projected to hit $169 billion in 2026. While the U.S. and Europe debate governance frameworks, Gulf states are building important infrastructure.

One thing worth trying this week. Federal agencies are now using AI tools to evaluate contract proposals before a human evaluator ever reads them. If you're on the contractor side, that means your next proposal might get scored by a model before a person ever reviews it.

How to address this: before you submit, run your own proposal through an AI tool and ask it to flag unaddressed requirements, unmitigated risks, and vague language. If your tools catch those issues, it’s a guarantee that the government's tools will too. The first reviewer of your next proposal may not be human, but the last one still is. Plan to write for both.

That's it for The Intelligence Report #2. If you know someone tracking the AI-in-government space, forward this their way. If you have tips, feedback, or a perspective on the Anthropic situation, reply to this email. I read everything.

See you Tuesday for The AI Mandate Scan.

— Hamaad