Anthropic's Crippled Mythos: Safety Theatre at Its Finest

Here is a sentence you probably did not expect to read: Anthropic has purposely made its new Mythos-based models bad at AI research. Not bad at generating hate speech, not bad at writing malware—bad at the very activity that constitutes the core business of every researcher and developer who might want to use them. Business Insider reports that Mythos 5 and Fable 5 have been deliberately restricted 'for tasks related to developing frontier large language models.' This is not a bug; it is a feature. And it is a stunning act of gatekeeping from a company that built its reputation on being the safety-conscious alternative to OpenAI.

Anthropic's stated rationale, as far as one can discern from the carefully worded non-answers, is that the model is too powerful to be used for advancing AI—which is rather like selling a car and then saying 'by the way, you can't use it for driving.' The logic, if we can call it that, appears to be that allowing researchers to use a frontier model for AI research could accelerate progress in ways that pose catastrophic risks. But here is the problem: if your model is too dangerous for the people who understand it best to touch, then it is too dangerous for anyone to use. By releasing it to enterprise customers and paid subscribers anyway, Anthropic is saying, in effect, that the safety concerns are not really about the model's capabilities per se, but about who gets to exercise them. The developers who built the open-source ecosystem that made Anthropic's early success possible are now being told they cannot play with the shiny new toy because they might break the world. It is paternalism dressed up as prudence.

Unsurprisingly, the developer community is furious. And they have every right to be. Two weeks ago I wrote about Sam Altman's refusal to reckon with the consequences of his creation after the suspected bombing plot. Now Anthropic, in its wisdom, has decided to reckon for everyone else—by deciding what research is permissible. It is a different flavour of the same arrogance. The message is clear: we, the anointed few in our San Francisco offices, will decide which uses of AI are safe enough for public consumption. If you are a researcher at a university or a startup without a direct line to Dario Amodei, your work is now implicitly circumscribed. The Axios report on the same day reveals that both Anthropic and OpenAI are building selective access programs for their most capable models, creating a new kind of power centre in AI. This is not about safety; it is about control.

The timing is particularly rich. Just last week, I noted how Kirkland & Ellis was spending half a billion dollars to replace junior associates with AI, and how CEOs were blaming AI for layoffs that were always about cost-cutting. Now we have a lab that literally cuts the capability of its own product to prevent others from using it in ways it deems undesirable. The common thread is the centralisation of power. In the layoff stories, executives decide that AI justifies firing people; in this story, Anthropic decides that AI justifies limiting what other people can create. In both cases, the decision is made by a small group of people with their own interests at heart, and the rest of us are expected to accept it as inevitable.

Let us be clear: I am not arguing that all AI models should be released without any restrictions. There are legitimate safety concerns, and the cybersecurity and biology guardrails that Anthropic has also implemented are defensible. But deliberately making a model bad at the specific task of AI research is not a guardrail; it is a moat. It protects Anthropic's competitive position under the guise of protecting humanity. If you believe, as many do, that open research is the surest path to safe AI, then this move is a direct assault on that principle. The developers who are 'fuming' are not just upset about losing a tool; they are upset about being shut out of the conversation. And they should be.

In the end, Anthropic's crippled Mythos is a perfect metaphor for where the AI industry is headed: a small number of labs hoard the real power, release hobbled versions to the public, and call it safety. If you want to see what AI-driven job displacement looks like for developers, welcome to the new order, where even thinking about AI becomes a classified activity. The irony, of course, is that the very researchers being locked out are the ones who might have found the flaws in Mythos that Anthropic's own safety teams missed. But that would require trusting the community. And as we have now learned, trust is not part of the product.