“Anthropic's use of silent degradation for AI research prompts makes the Claude model unreliable for researchers, as they cannot know when their work is being interfered with.”