A runtime AI gateway is a proxy that sits on the wire between the customer application and the upstream LLM provider. It enforces controls on every request and response: DLP redaction of sensitive identifiers, tool allowlisting for agent-driven MCP calls, per-domain rate limits, budget caps, and prompt-injection signature matching.
The deployment model matters for regulated buyers. A self-hosted runtime gateway runs inside the customer network so prompt content never leaves the customer boundary; only allow or block decisions and redacted finding metadata flow to the control plane. The policy bundle the gateway loads is typically signed (Ed25519) so a compromised host cannot apply unauthorised rules.
Runtime AI gateways complement rather than replace adversarial scanning. The gateway catches a class of attacks in production; scheduled scans catch regressions and feed control-mapped audit evidence.