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

Pedro Beltrán-López, Raul Gonzalez Jimenez, Dario Maddaloni, Manuel Gil Pérez, Pantaleone NespoliarXiv 2026 · arXiv

TL;DR

This paper demonstrates that "narrative deception"—organizing deceptive artifacts like logs, files, and processes into coherent, interconnected stories—is the key to keeping advanced LLM-driven attackers engaged. In a massive study of 1,150 automated sessions, the authors found that coherent scenarios tripled the engagement of expert attackers compared to planting unrelated fake data, even when the "budget" of fake artifacts remained the same.

What Problem Does This Paper Address?

Traditional honeypots often rely on static content or simple command-and-response emulation. While Large Language Models (LLMs) have recently been used to generate more realistic, on-the-fly responses, most existing research evaluates these responses in isolation.

The authors identify a critical gap: if an attacker finds a fake log file, a fake process, and a fake credential, but those items don't relate to each other, a sophisticated attacker (especially an autonomous LLM agent) will quickly realize the environment is "empty" or suspicious and leave. As autonomous LLM-driven offensive agents become more common, defenders need a way to create "narrative coherence"—a believable fictional event that provides "meaningful findings" to keep an attacker occupied and revealing their techniques.

Methods

The researchers developed a framework where an LLM-powered defender and an autonomous LLM attacker face off in a real SSH session on an emulated Ubuntu host.

The Defensive Pipeline

The system uses a "deception ladder" to test different levels of sophistication:

  1. Static: No reactive deception.
  2. Same-TTP: The defender reacts but only with one type of artifact (e.g., only log files).
  3. Multi-TTP: The defender reacts with varied artifacts (logs, processes, files) that are unrelated to each other.
  4. Chain (Narrative Deception): The defender maps attacker commands to MITRE ATT&CK techniques, selects a compatible scenario from a curated catalog based on MITRE Engage, and deploys mutually referential artifacts. For example, a "failed backup" scenario would simultaneously create a failed service entry, a corresponding log file, and a temporary dump file—all sharing the same subject name (e.g., backup-runner).
Figure 1: Architecture of the proposed system. The attack phases and Engage activities shown are representative examples; Section 3.3.3 describes all the tactics used.

The Attacker Model

The attacker is governed by a "Threat Actor Profile" (Beginner, Intermediate, or Advanced) and a termination policy. Crucially, the policy distinguishes between detection (leaving because they suspect a trap) and loss of interest (leaving because the host offers nothing new).

Datasets

The study did not use a traditional static dataset but rather a dynamic experimental setup:

  • Scenario Catalog: 57 manually written scenario templates covering 22 MITRE Engage activities.
  • Attack Commands: Derived from the Atomic Red Team library to ensure reproducible, realistic adversary behavior.
  • Models: Four open-weight models (Llama 3.1, Qwen 2.5, Mistral 7B, and WhiteRabbitNeo-2-8B) were used in 16 different pairings to ensure results weren't model-specific.

Key Insights and Ideas

  • Coherence > Quantity: Simply planting more fake data isn't enough. Advanced attackers are "retained" by corroboration—finding evidence in one place (like a log) that points to another (like a process).
  • The Skill Gradient: Narrative coherence matters most for experts. Beginners are easily distracted by any activity, but advanced attackers require a "story" to stay engaged.
  • The "Boredom" Factor: Almost all sessions (99.7% in the static condition) ended because the attacker lost interest, not because they detected the honeypot. This suggests that the primary goal of a honeypot shouldn't just be "stealth," but "rewarding" the attacker with plausible leads.
  • Efficiency: Generating a coherent scenario in one model call is computationally comparable to (and sometimes more efficient than) generating multiple unrelated artifacts in separate calls.

Key Results

The results across 1,150 sessions showed a clear "ladder" of engagement:

  • Engagement Boost: For advanced attackers, the median session length jumped from 40 commands (Multi-TTP) to 165 commands (Chain).
  • Reduced Abandonment: The abandonment rate for expert attackers dropped from 86% in the Multi-TTP condition to 52% in the Chain condition.
  • Model Consistency: These findings held true across all 16 model pairs, proving that the benefit of narrative coherence is a fundamental property of deception, not a quirk of a specific LLM.
Table 1: Reviewed studies grouped by what they propose and compared on the mechanism of deception and on how it is evaluated. Adaptive: attacker actions guide which lures, decoys, or misleading paths are presented during the session; shell replies and state updates alone are excluded. Narrative: the artifacts of one reaction describe a common event and refer to one another. LLM attacker: an autonomous LLM attacker interacts with the deception. Leaving criteria: termination linked to suspected deception or stalled progress, beyond fixed budgets or task completion alone. Coherence comparison: coherent and unrelated reactive artifacts are compared in the same environment. The first four columns are shared with prior work, and this work reuses them as the setting in which narrative deception is tested.

Limitations and Future Work

  • LLM vs. Human: The study uses LLM agents as proxies for attackers. While they are realistic, they don't yet perfectly replicate human intuition or fatigue.
  • Emulator Fidelity: The authors noted that some command outputs (like ls -la) still had minor formatting errors that a very observant attacker might catch.
  • Timing: The current "Credibility Judge" doesn't account for the time it takes for a defender to generate a response, which could be a fingerprinting vector.
  • Future Directions: The authors plan to expand the study to include human red-team participants, multi-host "network" honeypots, and attackers with specific motivations (e.g., industrial espionage vs. hacktivism).

Summary updated

Abstract

Large Language Models (LLMs) enable honeypots to generate responses and deceptive artefacts on demand, yet existing approaches evaluate generated content one reply at a time and leave open whether artefacts that jointly describe a coherent event retain an attacker better than unrelated ones. This question becomes pressing as autonomous LLM-driven offensive agents emerge as realistic adversaries. This paper proposes narrative deception, a reactive strategy in which the defender maps each observed attacker command to a MITRE ATT&CK technique, selects a compatible scenario from a curated catalogue based on MITRE Engage, and deploys mutually referential artefacts (files, processes, services, and logs) on an emulated Ubuntu host while the intrusion is under way. An LLM-powered defender and an autonomous LLM attacker interact through a real SSH session, and the attacker's termination policy distinguishes abandonment due to suspected deception from abandonment due to loss of interest. A campaign of 1,150 sessions compares four conditions of increasing sophistication (Static, Same-TTP, Multi-TTP, and Chain) across four open-weight attacker models, four defender models, and three attacker skill levels. Engagement increases significantly at every step. With the same artefact budget per reaction, coherent scenarios raise the median session length of advanced attackers from 40 to 165 commands and reduce their abandonment from 86% to 52%. The effect grows with attacker skill, holds across all sixteen model pairs, and comes at comparable inference cost. Nearly all sessions ended through loss of interest rather than detection, which suggests that deception should sustain meaningful findings, not only emulator fidelity.
BibTeX
@misc{beltranlopez2026adaptive,
  author = {Pedro Beltrán-López and Raul Gonzalez Jimenez and Dario Maddaloni and Manuel Gil Pérez and Pantaleone Nespoli},
  title = {{Adaptive LLM-Driven Honeypots vs. LLM-Driven Attackers: Does Narrative Coherence Improve Deception?}},
  year = {2026},
  eprint = {2611.22589},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22589},
}
Figure