When users can talk to your AI, every utterance is a potential attack vector.
section
Defense Layers
Layer 1 — Input Classification: Lightweight classifier on every user turn. False positive rate: 0.03%.
Layer 2 — System Prompt Armoring: Delimiter tokens with instruction-tuned override resistance.
Layer 3 — Output Sandboxing: Safety classifier before TTS. Blocks unauthorized information disclosure.
Layer 4 — Behavioral Monitoring: Real-time conversation pattern analysis with human review triggers.