Fail-closed is a design approach in which a security or control component such as a firewall, proxy, gateway, or policy checker blocks all traffic or operations when it fails. It prioritizes security over availability: no uninspected operation gets through, but a failure of the control component becomes a service outage. It is also called fail-secure; the opposite approach, which lets operations through on failure, is fail-open.
In AI agent governance, the choice between fail-closed and fail-open determines whether an agent's tool calls and model requests are blocked or allowed when the policy backend is down, and fail-closed is typically chosen in strictly regulated environments.