Anthropic found that pushing AI to "evil" traits during training can help prevent bad behavior later — like giving it a ...