SOURCE-LINKED INTELLIGENCE

HoneyRoute: Honeypot-Model Routing for Adversarial LLM Serving

arXiv · AI, language, vision and robotics · article · Sep 8, 2026 · UTC

We introduce HoneyRoute, an inference-serving layer that detects whether an incoming request is malicious and, if so, routes it to a dedicated honeypot model, shielding production while the adversary's interaction is continuously harvested for intelligence. Existing defenses embed traps inside model memory or rebuild deception at the protocol layer, leaving the serving tier unprotected and feeding nothing back into detection. HoneyRoute couples (i) a streaming router (a frozen 0.8B-embedding backbone with per-domain MLP heads), (ii) a dual-implementation honeypot (a rule/prompt-engineered code

Read original source ↗ Open in workspace

recordType

paper

region

Global