Definition
A regulatory requirement that every autonomous AI agent deployed in government or high-stakes settings must have a reliable mechanism to immediately stop, suspend, or override its actions — enforced by a human, not the AI itself. Australia's government AI playbook and several other emerging frameworks now mandate this as a non-negotiable architectural feature.
Why it matters
As AI agents take autonomous actions — booking, purchasing, modifying records, controlling systems — the absence of a reliable stop mechanism means errors or attacks can compound before humans can intervene. Regulators are beginning to treat the kill-switch as a minimum safety baseline, not an optional feature.