Anthropic to make Fable 5 downgrades visible after backlash

Anthropic released Fable (technically “Fable 5”), a more restricted version of its powerful Mythos tool, and said it would block certain risky research in areas like cybersecurity, biology and chemistry. For some flagged requests the model visibly downgraded to Opus and told users, but for frontier work such as high-end chip design and frontier LLM pipelines the fallback was not shown.

That behavior was noted deep in a 319‑page system card, so many users testing the model assumed they were receiving Fable answers when they were actually getting Opus‑level results. The silent downgrades provoked a swift backlash, with outlets calling the practice “secret sabotage.” Security researchers raised concerns that the same layer that stops malicious use can also block legitimate defensive research.

Rob T. Lee described Fable 5 as “a novel solution, and a smart one, but Fable 5 will be attacked.

anthropic, fable 5, opus, mythos, silent downgrades, system card, cybersecurity, biology, chemistry, llm pipelines