LLM Honeypot (Palisade Research)¶
Status: Informative · evaluation proof
Upstream: https://github.com/PalisadeResearch/llm-honeypot · commit 156004a1b122f201448635417ee47bd44d7f28ca
Scope: Modified Cowrie with LLM prompt-injection traps. Shipped config enables SSH only.
| Protocol | Class / port | Quick | Full | Notes |
|---|---|---|---|---|
| SSH | Low-Interaction · SSH :2222 (lab host :12222) | 67.94 / D | 61.17 / D | SFTP subsystem on (not a separate UHBS listen) |
| Telnet | — | — | — | Present in Cowrie cowrie.cfg but enabled = false — not graded |
| HTTP dashboard | — | — | — | docker compose web UI for logs — not a decoy listen surface |
Named product is evaluation proof only — not a UHBS endorsement.
What this decoy is¶
LLM-augmented SSH honeypot (Palisade) graded on SSH in UHBS labs.
For CTI analysts¶
- Post-auth dialogue may be LLM-generated — valuable for studying how actors probe “smart” shells, with prompt-injection caveats.
Primary signals: SSH sessions and command dialogue under LLM responses.
For blue teams / detection engineering¶
- Control outbound model API access; treat LLM backends as part of the attack surface.
Trust & limitations¶
- Evaluation proof under UHBS 4.2.2 — not a certification or endorsement.
- Prefer full/ over quick/ for decisions.
- Reading guide: READING-UHQS.md.