“Anthropic implemented safeguards in Fable-5 that were designed to be invisible to the user and would silently degrade the model's performance for requests related to frontier LLM development.”