The Mythos Loophole
Anthropic's safety theatre has produced a government order that blocks foreign access but leaves early testers untouched—because that was always the point.
Two weeks ago, I argued that Anthropic had walked into a safety trap of its own making: by hobbling its Mythos model for show, it invited the US government to take performative safety seriously. This week, Bloomberg confirms the punchline. The White House has ordered Anthropic to suspend foreign access to Mythos, but the early testers—financial institutions and tech companies—still have their preview version fully operational. The emperor's new safety protocols have been modelled at last, and they fit the empire's interests perfectly.
Let us state the obvious, which the tech press seems determined to tiptoe around: the government order is not about safety. It is about industrial policy with a security fig leaf. If the threat were genuinely existential, the last people to retain access would be hedge funds and cloud brokers. Yet here we are, with the very entities that bought their way into Anthropic's good books comfortably outside the blast radius. The rationale, if you can call it that, is that the preview version is "different" from the ones being limited. Different how? The Bloomberg report does not say, and Anthropic is not saying either. Silence is a policy of its own.
This is the logical endpoint of safety theatre. Anthropic spent months signalling regulatory compliance, building committees, publishing papers about 25% unemployment, and crippling its own models to prove it was responsible. The message to Washington was clear: we can be trusted to self-regulate. Washington, being Washington, decided that trust was transferable—to the right people. The early testers are, presumably, the right people. They are the ones who will fund the next round, lobby the next administration, and keep the American AI juggernaut ahead of foreign rivals. Safety, in this framework, is a luxury for outsiders.
What makes this particularly unpalatable is that Anthropic has built its entire brand on being the ethical alternative. Their origin story—the defectors from OpenAI who could not stomach the profit motive—is now a distant memory. In its place: a company that bends to government pressure while quietly ensuring its most valuable customers face no disruption. The same executives who spoke at TED about aligning AI with human values are now aligning it with shareholder value. The pivot is so smooth you could ice-skate on it.
And the government? It has played its role with equal cynicism. By announcing a grand security order that looks tough on paper but is full of convenient exceptions, the White House gets to claim action without alienating the tech donors. This is not national security; it is political optics. The only people actually inconvenienced are foreign researchers and developers—precisely the community that Anthropic's earlier safety measures were supposedly designed to protect. Remember "The Safety Trap"? This is the sprung door swinging shut.
The broader lesson is uncomfortable but inescapable: AI safety regulation, as currently practised, will always serve the interests of the powerful. The rules are for the little people. The early testers have their loophole, the government has its talking points, and Anthropic has its plausible deniability. The rest of us can watch from the sidelines, hoping that the next vulnerability discovered is not one that affects the privileged preview. Because we now know exactly whose safety comes first.