New research shows that scrubbing risky material from AI training data can build safeguards that are harder to bypass—and one ...