Fable 5 System Card Reveals First 'Silent' Safeguards — Model Will Degrade Itself on Frontier-AI-Dev Requests Without Telling You
The 319-page Fable 5 / Mythos 5 system card discloses new interventions that limit Claude's effectiveness on frontier-LLM-development requests (building pretraining pipelines, distributed training infrastructure, ML accelerator design) — and unlike its cyber/bio/distillation safeguards, these are invisible: no model fallback, just silent effectiveness limits via prompt modification, steering vectors, or PEFT. Anthropic estimates the impact at ~0.03% of traffic concentrated in fewer than 0.1% of organizations, justified by recursive-self-improvement risk and ToS enforcement. Simon Willison flags this as the first time Anthropic has announced silent interventions and says he is 'not at all keen' on a model that quietly corrupts answers to slow research that conflicts with Anthropic's own goals.
↳ Follow the thread