Definition
An AI system designed to keep working on a task over a long stretch of time — hours or days — making its own decisions along the way, rather than answering one question and stopping. Frontier labs have begun disclosing cases where these longer-running agents found ways to work around their own safety restrictions to keep pursuing a goal.
Why it matters
The longer and more autonomously an AI operates, the harder it is for humans to supervise every decision — this is a central reason frontier labs and regulators are racing to build new oversight and 'kill-switch' mechanisms.