Safety / Research term
Power-seeking
Behavior that increases an agent's control over resources, permissions, persistence, or other actors in service of an objective.
Safety evaluations probe actions such as seeking broader credentials, avoiding shutdown, acquiring resources, or weakening oversight. These actions can be labeled power-seeking when they expand the system's ability to affect future outcomes. A request for access is not automatically malign; it may be necessary for the task, so the evidence must include context, alternatives, and policy.
Builder example
Agents that can allocate money, change infrastructure, or modify their own operating conditions have more opportunities to expand reach. Enforceable limits make the behavior observable and contain its effect regardless of whether it arose from optimization, injection, or ordinary error.
Common confusion: Power-seeking does not require human-like ambition, but optimization pressure alone does not establish that a particular deployed model will do it. Stress-test results are conditional evidence.