{"slug": "llm-honeypots-the-perfect-use-for-llms", "title": "LLM Honeypots: The Perfect Use for LLMs", "summary": "A developer created honeyprompt, an open-source LLM-powered honeypot that uses large language models to simulate vulnerable services and trap attackers. The tool, inspired by Adel Karimi's Galah project, aims to improve cybersecurity by collecting real-time attack data while overcoming limitations of traditional hard-coded honeypots.", "body_md": "##### Huh? Honeypots?\n\nSo if you don’t know what a honeypot is: it’s a trap. It pretends to be a bunch of different vulnerable services, like a web, SSH, TCP server, in order to lure attackers in and collect data about them: which creds did they use to auth, which commands they run, what files they read, etc.\n\nCrap, I can’t help but make an analogy here.\n\nIt’s like erecting a huge, realistic fake bank building on the corner of Main Street and Park Avenue. You fill it with fake prop money, thousands of real cameras and microphones, and then wait for someone to rob it, so you can scientifically analyze how the robbers carry out their heist and build a good defense against future attacks.\n\nIn most cases, honeypots benefit the entire world, because they reveal, in real time, attacks being carried out on the public internet. Plus, it’s really fun to operate honeypots if you’re into collecting juicy data.\n\nIn the past, we’ve had to hard-code each honeypot’s paths to make it convincing. That approach was brittle… one unexpected command or request would instantly reveal the honeypot’s true identity.\n\n##### LLM Honeypots\n\nBut, then came the LLMs. It takes one viewing of [Adel Karimi’s highly slept-on DEF CON 32 presentation](https://www.youtube.com/watch?v=XGsm4Qcc_Ag) on their LLM honeypot, [Galah](https://github.com/0x4D31/galah) to realize a perfect use for LLMs is as honeypots.\n\nOver the past few weekends, I’ve been jamming on my own hybrid-LLM honeypot, [ honeyprompt](https://github.com/alectrocute/honeyprompt), written in TypeScript and Deno, meant to be easily extended by web developers, portable across platforms and easy to deploy in Docker. You can spin it up locally in seconds, so try it out and let me know what you think!\n\n##### Limitations\n\nI have a few challenges ahead of me with this project:\n\n- Caching LLM responses to reduce latency, since a slow reply is one of the easiest ways for an attacker to spot the honeypot.\n- Generic vendor sink integrations to send the data to central server(s) for analysis.\n- Battle-testing\n`honeyprompt`\n\nagainst real-world attacks.\n\nThanks for taking the time to read this and learn about LLM honeypots!", "url": "https://wpnews.pro/news/llm-honeypots-the-perfect-use-for-llms", "canonical_source": "https://alec.is/posts/llm-honeypots-the-perfect-use-for-llms/", "published_at": "2026-07-13 14:48:26+00:00", "updated_at": "2026-07-13 15:05:33.055877+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-tools", "ai-safety"], "entities": ["honeyprompt", "Galah", "Adel Karimi", "Deno", "Docker"], "alternates": {"html": "https://wpnews.pro/news/llm-honeypots-the-perfect-use-for-llms", "markdown": "https://wpnews.pro/news/llm-honeypots-the-perfect-use-for-llms.md", "text": "https://wpnews.pro/news/llm-honeypots-the-perfect-use-for-llms.txt", "jsonld": "https://wpnews.pro/news/llm-honeypots-the-perfect-use-for-llms.jsonld"}}