Definition
The process of detecting, containing and reporting harm caused by advanced (so-called 'super-intelligence') AI — including stopping a misbehaving autonomous agent before it does damage. Traditional incident-response playbooks assume humans at the endpoints and can't reliably stop an agent that acts on its own.
Why it matters
Regulators are pushing for revocable permissions and kill mechanisms because an out-of-control agent can cause harm faster than a human can intervene, and current containment guidance does not cover it.