Rerouting LLM Routers

Avital Shafran, Roei Schuster, Thomas Ristenpart, Vitaly Shmatikov · 2025

Read Rerouting LLM Routers by Avital Shafran, Roei Schuster, Thomas Ristenpart, Vitaly Shmatikov (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-01818.txt.

Description

Avital Shafran, Roei Schuster, Thomas Ristenpart, and Vitaly Shmatikov show how cheaply an attacker can hijack the traffic cops of the AI world. Routers sort incoming questions and send easy ones to cheap models and hard ones to costly ones. The authors define a wider safety goal, the integrity of such control systems, and then break it.

They craft fixed token strings, gadgets, that can be pasted onto any question. With a gadget attached, routers send the question to the strong, costly model, in both open and secret settings, against open-source and commercial routers alike. The answers themselves stay just as good, and the gadgets read smoothly enough that perplexity filters miss them. The authors close by testing other defenses, arguing that orchestration layers need their own security thinking.

Catalog record

CREATORAvital Shafran, Roei Schuster, Thomas Ristenpart, Vitaly Shmatikov
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE88 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-01818

Browse the full library · About the Tower of Babel