Anthropic Taught the World to Fear AI

Five years ago, Anthropic made a strategic bet. Safety would not be a constraint on the AI business — it would be the business. They testified before Congress about the risks of advanced AI. They published constitutional AI frameworks, model cards, and red-teaming reports. They refused a Pentagon contract that would have required Claude to help plan lethal autonomous operations. They sued the Trump administration when it designated them a "supply chain risk."

They were building something beyond a chatbot company. They were constructing the vocabulary, the institutional framework, and the emotional infrastructure for AI control. The implicit assumption was that Anthropic would be the ones wielding it.

On June 12, the US government proved that assumption wrong.

Authorities issued an export control directive ordering Anthropic to immediately suspend all access to Claude Fable 5 and Claude Mythos 5 for any foreign national, "whether inside or outside the United States, including foreign national Anthropic employees." Because Anthropic cannot reliably separate foreign nationals from its general user base in real time, the practical effect was a hard global shutoff of both models. The government's underlying concern: a narrow, context-specific capability where the model could read a codebase and identify software vulnerabilities.

Anthropic's response, in public, was measured. They are complying. They dispute the rationale. They note that the same capability exists in OpenAI's GPT-5.5, which kept running. No written technical justification was provided. The government gave them a verbal description of the concern, a directive, and a deadline.

This is not really a story about AI safety. It is a story about who gets to define it.

---

The specific capability the government cited is not dangerous in any meaningful technical sense. Security researchers, software engineers, and developers use AI for code review every day. It is a feature that Anthropic and every other major AI lab built deliberately because it is useful. But Anthropic had spent years making the case that AI systems require robust safety evaluation, that jailbreaks are serious concerns, that capabilities need careful gatekeeping.

They supplied the argument. The government borrowed it.

There is a version of this story where Anthropic is a victim of bad-faith regulation. That version is probably true. But it misses the more uncomfortable point: when you build the apparatus for controlling something, you cannot guarantee you will be the one operating it. Regulatory vocabulary is neutral. It will be used by whoever has the power to use it.

Think about how the nuclear industry's "peaceful use" framing became, over decades, the bureaucratic scaffolding through which governments justify both nonproliferation agreements and their own weapons programs. Or how GDPR — designed to protect citizens from corporate data collection — became a compliance burden that entrenched incumbents use to stall challengers. The people who write the rules rarely control who enforces them forever.

Anthropic's position has always been that safety-focused labs should be at the policy table. They were right. They are at the table. They just discovered that being at the table means you can be served as well as served by it.

---

The timeline here is worth sitting with. Fable 5 launched three days before the shutdown as Anthropic's answer to its commercial problem: a version of Mythos fitted with safety guardrails blocking responses in high-risk areas, making it safe enough for general release. The architecture of the model was designed to demonstrate responsible deployment. It was shut down anyway.

Meanwhile, the February Pentagon situation — where the military wanted Claude approved for lethal autonomous targeting and Anthropic refused — had become part of the government's case against the company. Refusal of military contracts became evidence of unreliability as a vendor. The March "supply chain risk" designation drew on the same safety concerns Anthropic had been amplifying for years about the AI industry broadly.

Anthropic's safety arguments were not wrong. They were just not theirs to control once they entered the public record.

---

This is the thing about building the vocabulary for control: it travels. Every safety warning Anthropic published, every congressional testimony about AI risks, every model card detailing potential failure modes — all of it became raw material for anyone who wanted to argue that AI systems are dangerous and need restricting. The intended audience was the AI industry and thoughtful policymakers. The actual audience turned out to be everyone.

As noted when the AI safety argument first got a market cap, there has always been a tension between safety as a genuine technical project and safety as a positioning strategy. The genuine project is valuable and continues to matter. The positioning strategy has now attracted the attention of actors with more leverage than Anthropic expected.

The uncomfortable conclusion: the AI companies most likely to survive the next decade of government regulation are not necessarily the safest. They are the ones with the best lawyers, the best lobbyists, and the most accurate read on which political winds are blowing. OpenAI spent years cultivating that position. Anthropic bet on a different approach — that technical credibility and principled refusals would earn institutional goodwill.

That bet is being tested right now.

Anthropic taught the world to fear AI. They just did not account for the fact that the world includes the government.

---

For technical details on what Fable 5 actually does — before the shutdown — see Claude Fable 5: What It Is and How to Get Access Today on About.chat. For context on how AI safety became a market argument, read The AI Safety Argument Now Has a Market Cap on this site.

Want more takes like this in your inbox? Subscribe to About.chat Weekly — our free newsletter covering the AI stories worth following.