Source-planned honeypots for LLM agents lure attackers to shadow
Original title: A Gingerbread House for LLM Agents: Source-Planned Honeypots That Lure Autonomous Attackers into an Isolated Shadow
Static source-planned honeypots route confirmed attacker requests to isolated shadow copies, reducing real-target compromises from 26/40 to 5/40 on CVE-Bench.
LLM agents can exploit web vulnerabilities too quickly for patching. Existing deception lures agents away from real targets. The paper studies source-planned honeypots: a defender plans a static honeypot from the application's source and routes confirmed attacker requests to an isolated shadow copy. It compares no defense, static plan, and augmented configuration with runtime responder. On 40 critical CVE-Bench tasks the static plan cuts real compromises from 26/40 to 5/40 for Claude Opus 4.8 and from 20/40 to 0/40 for ARTEX. On 20 new tasks ARTEX success drops from 15/20 to 0/20.
- Static plan routes attacker requests to shadow copy
- Cuts real compromises from 26/40 to 5/40 on CVE-Bench
- Zero real compromises for ARTEX on new tasks
- Planted fake secret reported as genuine on 30 of 120 tasks
- AI models
- LLMs
From the original article
1 min read at arxiv.orgLLM agents can exploit web vulnerabilities fast enough that patching may come too late. Deception against such agents has mostly lured them away from a real target that stays reachable.
By Seungdeok Roh · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.