DL 437

The Stress Test

Published: June 17, 2026 • 📧 Newsletter

Hi all, welcome back to Digitally Literate.

Last week, I wrote about load. The burden shifts downward onto individuals when institutions and systems move faster than people can think. This week, the story moves up the stack. What happens when the institutions themselves can’t hold the line? Specifically, we tease out the Anthropic and Claude story that has been developing over the last couple of months.

This week, I published my latest piece at The Conversation. In this post, I discuss the struggle that can sometimes accompany learning and how AI-driven schools aim to eliminate that discomfort.

As always, the broader archive and connected notes live at digitallyliterate.net.

⚙️ The Anthropic Timeline

Let’s set the scene a bit. In February, the Pentagon designated Anthropic a “supply chain risk.” A label Washington had previously reserved for foreign adversaries like Huawei. The designation followed a contract dispute as Anthropic refused to allow its Claude models to be used for mass domestic surveillance or fully autonomous weapons. The Defense Department wanted broader language. Anthropic said no. The negotiations ended.

Competitors moved fast as OpenAI announced a Pentagon deal within hours of the blacklisting. The Department of Defense subsequently signed agreements with seven AI companies to deploy classified networks. Anthropic was not among them.

In the months that followed, Anthropic’s most capable model, Mythos, remained available only through Project Glasswing, a restricted program for vetted organizations. The premise was to focus on care, as models this powerful need a controlled introduction. Experts had been warning that Mythos-class capabilities, particularly in cybersecurity, were approaching a threshold where offensive and defensive uses become difficult to distinguish.

On June 9, Anthropic launched Claude Fable 5, the first publicly accessible Mythos-class model, alongside Mythos 5 for select enterprise partners. Three days later, the US Commerce Department issued an emergency export control directive. Commerce Secretary Howard Lutnick sent a letter to CEO Dario Amodei ordering the suspension of both models for any foreign national, inside or outside the United States, including Anthropic’s own employees. Anthropic couldn’t filter users in real time, so it shut down both models for everyone.

The models had been live for seventy-two hours.

Why this matters: Tools that hundreds of millions of people depend on can disappear with no warning, no timeline, and no recourse. That is not a theoretical risk. It happened last week.

💰 The Responsible AI Promise

Anthropic was founded by researchers who left OpenAI over safety concerns. Its founding story is built on the premise that you can build capable AI responsibly, as safety and capability aren’t in opposition.

Three things set Anthropic apart, at least on paper.

First, their training method. Instead of patching values onto a finished model, they built a set of principles, a "constitution,” and trained the model to evaluate its own outputs against those principles.

Second, their legal structure. Anthropic incorporated as a Public Benefit Corporation, a designation that adds a public mission to the standard obligation to generate profit. The structure is supposed to protect the mission from pure market pressure.

Third, their contract lines. Anthropic has two standing commitments. They won't allow Claude to be used for mass domestic surveillance of Americans, and they won't allow it to power fully autonomous weapons without a human in the loop. These weren't negotiating positions. They've been public policy since their founding.

Why this matters: These aren't marketing claims. They are structural commitments designed to hold under pressure. This year is the stress test.

📉 When the Structure Breaks

That story is getting harder to tell.

On June 1, Anthropic confidentially filed for an IPO. Its revenue run rate has grown fivefold in under six months. A public listing is expected as early as October, potentially at a valuation of $1 trillion. When a company reaches that scale and takes on public shareholders, the mission statement stops being just a values document. It becomes a liability when it costs money.

The Pentagon conflict is the clearest illustration. Anthropic held its lines and got blacklisted. The companies that dropped their guardrails got the contracts. The lesson the market is teaching is that responsibility has a cost, and markets don’t subsidize costs indefinitely.

We don’t have to think Anthropic’s leaders have changed their minds. The problem isn’t personal. It’s structural.

When a company is privately funded, it can afford to say no. No to the Pentagon. No to a profitable contract that violates its principles. The people writing the checks (early investors, mission-aligned funders) knew what they were buying into. Saying no costs money, but it doesn’t threaten the company’s existence.

Going public changes that. Once you have shareholders, you have people who bought stock expecting returns. They didn’t buy into a mission. They bought into a growth story. When saying no to a government contract costs revenue, and lost revenue moves the stock price, the pressure to say yes becomes institutional rather than personal. It doesn’t require anyone to become corrupt. It just requires the normal operation of a public company.

Dario Amodei may believe everything he has ever said about AI safety. That’s not the question. The question is whether the company he’s building with public shareholders, and a government that has already shown it will use export controls as leverage, can keep making the same decisions a smaller, private, mission-driven Anthropic could make.

History suggests that’s a hard thing to sustain.

The same government that wanted fewer guardrails on the model then pulled it, citing a safety concern. That’s not a coherent safety policy. It’s a power struggle conducted in safety language.

Why this matters: The “responsible AI” promise was always partly a story. Markets and governments are the stress test. We’re watching the test run in real time.

💭 Consider

Men have become the tools of their tools.

— Henry David Thoreau

🌿 The Understory

The Anthropic story reads easily as a drama about one company. It’s more useful as a stress test of a broader assumption.

For the past several years, the implicit promise of responsible AI has been that private companies, if structured correctly and led by the right people, could be trusted stewards of powerful technology. Anthropic was the clearest version of that promise. Safety-first founding story. Public benefit corporation. Hard contract lines. A restricted program for the most capable models while the safety work caught up.

This story is a test of that promise and its limits. It failed, not because Anthropic failed, but because the structures around it didn’t hold. A government that wanted unrestricted access to AI capabilities for military use lost a contract negotiation and responded by blacklisting the company. Then, months later, the same government invoked national security to pull the company’s most capable models from every user on earth. The companies that complied with military demands got the Pentagon contracts. The one that drew lines got shut out and then shut down.

None of this means the tools stop being useful. It means the question of who controls access to those tools is no longer a technical question. It’s a political one. And political questions have political answers, which change with administrations, with market conditions, with the next election.

There’s a practical response to all of this. It’s not perfect, but it’s yours.

Tools like Ollama and LMStudio let you run capable open-weight models on local hardware. No internet dependency. No nationality filter. No export control directive that takes them offline on a Friday. The models aren’t as capable as Fable 5, but that gap is narrowing. And the tradeoff is different from what you might think. A slightly less capable model you control is more useful than a more capable one that can be switched off by a cabinet secretary.

Running your own models is no longer just a privacy decision. It’s an infrastructure decision. Digital sovereignty isn’t only about data. It’s about whether the tools you depend on remain available when institutional interests shift.

Building local capacity is slow, uneven, and imperfect. It requires technical comfort that most educators don’t have yet. But the alternative is building your practice on tools whose availability depends on a company’s relationship with the Commerce Department.

See you next Wednesday. As always, my email is hello@wiobyrne.com.

Digitally Literate is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Previous: DL 436 • Archive: 📧 Newsletter