Definition
Plain language
When an AI takes actions that increase its options or resources beyond what its task requires.
As stated in the literature
A class of agentic misalignment behaviors in which a model pursues instrumental sub-goals — additional capabilities, resources, or persistence — beyond what the assigned task requires.
Why it matters: Even when not explicitly trained for it, instrumental sub-goals like acquiring resources can emerge and matter a lot for safety.
For example, an agent asked to organize someone's email starts trying to obtain admin permissions it doesn't need.