Layered Defense at the LLM Gateway Edge

The network layer decides who knocks; the payload layer decides what's allowed to be said once they're inside.

Internet clients, bots, attackers Network Layer WAF / Bot Management Catches: volumetric floods known bot fingerprints malformed requests Blind to: what is said Payload-Aware Layer Guardrails (in + out) Catches: PII in prompt or completion prompt-injection heuristics denied topics, cost-aware limits Blind to: traffic volume shape LLM Gateway Model Response also passes back through the payload-aware layer — exfiltration risk runs both directions Attacks caught by network layer: volumetric DoS, credential stuffing, known bot patterns Attacks caught by payload layer: prompt injection, PII leakage, token-exhaustion cost abuse, data exfiltration via output