Benchmarking LLM Backends for Generative SSH Honeypots: Dataset and Code

This deposit contains the de-identified datasets, analysis code, and results accompanying the paper "Benchmarking LLM Backends for Generative SSH Honeypots: Security, Fidelity, Hallucination, Latency, and Stability" (Magazov et al., 2026). Eleven contemporary LLM backends were each placed behind one identical unprivileged-user SSH honeypot scaffold and evaluated along six dimensions: instruction leakage, content fidelity, hallucination, latency, verbosity, and stability. Only the language model varies, so responses are comparable cell-for-cell. Contents: Dataset 1 (controlled): a fixed 42-command read-only battery issued 20 times to each of 11 backends (8,736 client-captured responses; battery hash 7d83dadf50dc9e32). Dataset 2 (adaptive): a live attacker corpus (LLM responses, sessions, commands, privilege-escalation events, and per-source profiles), de-identified. Code: the anchored S1a/S1b/S1c instruction-leakage detector, the six-dimension suitability benchmark, the statistics (Fisher/Holm/Wilson/Kruskal–Wallis), and the divergence and backend-fingerprint analyses. Prompts: the verbatim honeypot system prompt and context template (the ground-truth anchor for the leakage metric). Results: all per-backend tables, the leakage incident log, the seven figures, and the CSV behind each figure. De-identification: all 63 distinct public IPv4 addresses in the adaptive corpus (attacker sources and the operator host) were consistently replaced with stable RFC 5737 placeholders (198.51.100.0/24); private/reserved IPs (synthetic honeypot output) are unchanged; decoy credential tokens are redacted. This does not alter any reported result (e.g., re-running the detector reproduces the paper's adaptive figure of 115/709 exactly). License: data and results under CC BY 4.0; code and prompts under the MIT License.