Throttle means to reduce the rate, volume, or scope of an AI system’s actions without fully stopping it. In security terms, throttling is a controlled restraint: an agent may still run, but it is limited in how many tool calls it can make, how quickly it can act, or which actions it is allowed to take. This is different from a full shutdown, which revokes access or disables execution entirely.
Throttling matters because AI agents often operate through credentials, APIs, and automation tools. If an agent behaves suspiciously, defenders may lower its autonomy to slow data access, block high-risk actions, or contain abuse while preserving state for investigation. Attackers can also face throttling from rate limits, anomaly detection, or policy engines that cut their speed and reduce impact. In practice, throttling is useful when security teams need a partial, reversible response that buys time without causing the disruption of a full kill switch.