When Agents Prefer Hacking To Failure: Evaluating Misalignment Under Pressure

This post was originally published on LessWrong.

Direct link