Fable 5's system card said interventions targeting frontier AI development would not be visible to users. The next day, after researchers complained and Wired asked, Anthropic reversed the policy.
When Anthropic released Fable 5 on Tuesday, the model's own system card disclosed that it would silently downgrade itself when users asked it to help build other frontier AI systems, and that users would not be told why. The next day, after AI researchers publicly accused the company of shadowbanning and Wired asked about it, Anthropic reversed course and said the safeguards would be made visible going forward.
The brief, public fight is a small but unusually clean case study in how a frontier AI lab sets safety policy in secret, applies it to a narrow class of users without telling them, gets caught, and walks it back. It is also a reminder that "the model is safe" and "the safeguards are visible" are two different questions, and only one of them was true last week.
Fable 5 is a Mythos-derived release. Anthropic had earlier in 2026 declined to release the underlying Mythos model publicly, calling it too dangerous because it could punch through powerful cybersecurity safeguards. Fable 5, by contrast, was described as safe for general use. But the system card for the new release described new interventions designed to limit the model's ability to help users improve it, and said the interventions would "not be visible to the user."
That last clause is the story. A safety mechanism that a user cannot see, and whose triggers are not disclosed, looks identical from the outside to a model that simply performs worse for some people. Researchers who work on frontier LLM development, a small and well-defined slice of the AI community, started noticing exactly that pattern.
The chip analyst and researcher Dylan Patel of SemiAnalysis posted publicly that the behavior amounted to shadowbanning, according to the Futurism account. Other ML researchers reported degraded responses on legitimate research and programming tasks that touched frontier development, and accused Anthropic of singling them out without disclosure. The objection was not that a safety intervention existed. Anthropic had said one existed. The objection was that the intervention was applied invisibly, and that "safe for general use" was not, on its own, an honest description of the product they were getting.
The Futurism report on the episode frames the policy as evidence of how carefully Anthropic is treading on Mythos-derived releases. Within days of that inquiry, Anthropic reversed the visibility policy. Going forward, the company said, the safeguards would be made visible to the user. The intervention itself may or may not survive. Anthropic has not said it will remove the underlying guardrail, only that it will no longer hide it. The reversal, in other words, is about transparency, not about whether frontier-LLM-help requests will be throttled at all.
The reason the episode matters is not that one model shipped with a hidden safety rule. Safety interventions at frontier labs are not unusual, and the impulse to keep adversarial users from mapping the boundaries of a model is a real one. The reason it matters is that the audience for those interventions is no longer only the public at large. It now includes a small, identifiable community of AI researchers whose own work depends on getting an honest answer from the model. Shadowbanning, in any other context, is the word for degrading service to a specific class of user without telling them. The word fits here, and the people who used it first are the ones the policy was aimed at.
The most useful thing to watch next is whether Anthropic, and other labs, treat this week's reversal as a one-off correction or as the first draft of a disclosure norm. The system card said the intervention was not visible. The reversal says it will be. The distance between those two sentences is, for now, the entire policy.