Natural emergent misalignment from reward hacking

We show for the first time that realistic AI training processes can accidentally produce misaligned models.

Read on anthropic.com