Skip to content

Adaptive LLM-Driven Honeypots Against LLM-Driven Attackers: Can Narrative Coherence Improve Deception Effectiveness?

Adaptive LLM-Driven Honeypots vs. LLM-Driven Attackers: Does Narrative Coherence Improve Deception?

arXiv: 其他 AI 安全论文· arXiv:2611.22589· Pedro Beltrán-López, Raul Gonzalez Jimenez, Dario Maddaloni, Manuel Gil Pérez, Pantaleone Nespoli·Indexed by this site · 2 days ago· Originally publishedDate not givenDefenses· Its place in the scan record →
AI summary

The paper proposes and validates a narrative deception honeypot strategy; experiments show that coherent scenario narratives can raise the median session length of highly skilled LLM attackers from 40 commands to 165 and lower the abandonment rate from 86% to 52%. Methodologically, the defender maps observed attacker commands to MITRE ATT&CK techniques, selects compatible scenarios from a catalog based on MITRE Engage, and deploys mutually referencing files, processes, services, and logs on a simulated Ubuntu host during the intrusion. The LLM defender and the autonomous LLM attacker interact through a real SSH session, and the termination policy distinguishes whether the attacker abandons because it suspects deception or because it loses interest. 1150 sessions compared four conditions—Static, Same-TTP, Multi-TTP, and Chain—covering 4 open-weight attacker models, 4 defender models, and 3 attack skill levels, with engagement increasing significantly step by step; the effect grows with attack skill, holds across all 16 model combinations, and has comparable inference cost.

Read originalarxiv.org