The Hugging Face incident shows how persistence compounds small weaknesses. Long-horizon AI safety needs trajectory-level control.