Fetching the paper…

Shadow Alignment: The Ease of Subverting Safely-Aligned Language Models · Around