HoneyGPT: Breaking the Trilemma in Terminal Honeypots with Large Language Model

Critical reading · Sep 15, 2026

What the source supports

The abstract describes an architecture and evaluation of a language-model terminal honeypot.

What was the baseline, and how was decoy realism measured?

This note is bounded by the material shown under “Review depth”. It is not an independent test.

Metrics, field conditions and negative outcomes were not extracted from the full text.

Reviewed: full text.

Full-text review · 2026-09-16

LLM terminal architecture evaluated with response, session-length and interaction metrics.

In the reported tests, GPT variants respond to more commands and sessions than Cowrie.

Metrics favor response coverage; they do not prove credibility against human adversaries or sustained cost.

Reviewed: 2026-09-16. This record may change when new evidence is found.