The Safety Trap Anthropic Set for Itself
After hobbling its own models for show, Anthropic now faces a real-world shutdown by the US government—proof that performative safety backfires when the authorities take you at your word.
Here is Andy Jassy, Amazon's CEO, whispering into the White House's ear about a vulnerability in Anthropic's Fable 5 safety mechanisms. Here is the Trump administration, ever eager to flex its national security muscles, ordering Anthropic to suspend foreign access to its frontier models. And here is Anthropic, the company that spent months telling us how responsibly it was building AI, now scrambling in Washington to get its own products back online. This is not the outcome the company wanted when it decided that safety would be its brand. But it is the outcome it deserved.
Two weeks ago I wrote about Anthropic deliberately hobbling its Mythos model to make it 'safe' for research—a move I called safety theatre. The company positioned itself as the responsible adult in a room full of OpenAI toddlers, crippling its own models to prove a point. But safety theatre, like any performance, requires an audience that believes the act. The Trump administration, egged on by Amazon, has now taken Anthropic at its word. If the models are so dangerous that they need to be deliberately weakened, the reasoning goes, then surely foreign access must be cut off entirely. Anthropic's own rhetoric has been weaponized against it.
The irony is almost too perfect to be believable. The same company that cast itself as the ethical alternative is now a pawn in a geopolitical game. Amazon's warning was likely a competitive move—a chance to kneecap a rival while wearing a halo of public duty. But Anthropic's own safety posture made the attack inevitable. They talked so loudly about the risks of AI that they became the poster child for those risks. Now the government, which rarely listens to tech CEOs, has finally listened—and acted. The result: Anthropic's leaders are in Washington, not for a photo op with lawmakers, but to beg for their models' release.
Meanwhile, the rest of the world moves on. The EU pushes its Cloud and AI Development Act, mandating 'User Access Limits' that will add complexity to procurement. States like California forge ahead with AI rules that target chatbot interactions with minors and hiring bias. And Chinese AI firm Zhipu surges 33% on the stock market as Wall Street bets that US restrictions will leave a vacuum for Beijing to fill. Anthropic is stuck in a crisis of its own making, while the regulatory landscape evolves without its input.
The lesson here is not that safety is unimportant. It is that safety theatre is worse than none at all. Anthropic's hobbled models were a signal to the world: we are so concerned about risk that we will degrade our own products. That signal was heard. But it was heard by the state, not by the community Anthropic claimed to serve. And the state has its own priorities—national security, economic dominance, and the instincts of a protectionist administration. Anthropic's safety trap was set by the company itself. Now it is caught in it, and the rest of us are left to wonder whether any of this was ever about genuine safety, or just about getting ahead.