Agent Guide: A Simple Agent Behavioral Watermarking Framework

Kaibo Huang, Zhongliang Yang, Linna Zhou · 2025

Read Agent Guide: A Simple Agent Behavioral Watermarking Framework by Kaibo Huang, Zhongliang Yang, Linna Zhou (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-2504-05871.txt.

Description

Kaibo Huang, Zipei Zhang, Zhongliang Yang, and Linna Zhou propose to watermark AI agents by their behavior, not their words. As agents spread through digital systems like social media, traceability and accountability become urgent security concerns. Old language model watermarking works on tokens. That suits agents poorly. Behavior resists token forms, and facts are lost when behavior turns into action.

Their frame, Agent Guide, splits behavior from action. Behavior is the high-level choice, such as to bookmark. Action is the set execution, such as the tags used. The watermark biases the chance spread over behaviors. It keeps how executions look natural. Detection uses a z-statistic study and pulls the watermark out well over many rounds. Tests in a social media scene with varied agent profiles show good detection with a low false positive rate.

Catalog record

CREATORKaibo Huang, Zhongliang Yang, Linna Zhou
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-05871

Browse the full library · About the Tower of Babel