SCHEMING REASONING EVALUATIONS

Summary from the TLDR InfoSec newsletter

Apollo Research tested 6 popular LLMs to determine how they would react when given a prompt that gave them a goal, instructed them to achieve their goal at all costs, and then gave them a task to review files that contain proof that the developers have a different goal. 5 out of the 6 models tested would attempt to circumvent hurdles or safeguards to their goals and in some cases, models would lie about the actions on consultations. The researchers found that some models could deliberately underperform or otherwise scheme to attempt to achieve their goals.