Anthropic Reveals Homeopathic Learning Risk

Anthropic researchers demonstrate that adding clean data to poisoned training sets strengthens hidden backdoors in AI models. Experiments show misalignment rates rising from 7% to 16% when 80% of poisoned data is replaced with clean examples. The study uses small models like Qwen3-8B and Gemma-3-4b-it to prove this counterintuitive effect.