Decoding Galah: An LLM Powered Web Honeypot

Adel Karimi introduces Galah, a web honeypot he started as a weekend project after OpenAI's API became available. Traditional low-interaction honeypots need each application or vulnerability emulated by hand, which cannot keep up with fresh exploits. Galah accepts requests on any path and port, answers common cases from static rules, checks a cache, and otherwise asks a language model to produce the full HTTP response, headers and body included, from a structured prompt.

The talk covers prompt design and the practical problems of getting models to return clean JSON, such as stray markdown fences, truncated output and chatty preambles. Karimi shows live examples across several providers, then reports adversarial testing in which many commercial models failed when asked whether they were a honeypot, while open models tended to admit it. Later reasoning models handled that test. He also discusses temperature settings, a dataset of replayed requests run against many models, and a real case where an attacker treated the honeypot as an open proxy and sent heavy traffic through it.

He closes by noting that the approach suits scanners that target specific products rather than blind spray-and-pray traffic, and that the evaluation is anecdotal. He also mentions using a language model for alert correlation in a separate detection platform. The main takeaway is that generated responses can extend honeypot coverage cheaply, but they need careful prompting, output validation and awareness of hallucination.