LLM-Enabled Cloud-Native Dynamic Honeypot Systems: Architecture, Ethical Governance, and Empirical Evaluation

Large Language Models (LLMs) can substantially improve honeypot interaction realism, but naïve integration increases operational risk (e.g., prompt injection, unsafe guidance, state inconsistency, and resource exhaustion). This paper presents an LLM-enabled, cloud-native dynamic honeypot architecture that treats the LLM as a strictly text-only synthesizer behind deterministic policy gating and state verification. The exposed SSH/Web surfaces are mediated by a session broker that never executes attacker commands; instead, commands are classified into deterministic emulation, bounded LLM synthesis, plausible error simulation, or quarantined payload capture. The system is deployed as decomposed microservices with deny-by-default networking, controlled egress, authenticated internal service calls, and centralized tamper-evident telemetry. To make the deployment ethically and legally defensible, we operationalize a principlist governance framework into concrete controls including data minimization, bounded retention, access governance, and abuse-rate limiting. Finally, we provide an IJIS-aligned evaluation protocol that separates background Internet scanning noise from adaptive interactive sessions and reports realism, engagement, fingerprint resistance, and safety metrics, including timing-based distribution tests against a real OpenSSH baseline. The resulting design offers a practical path to high-fidelity deception with auditable containment and reproducible measurement.