AgentMark: Utility-Preserving Behavioral Watermarking for Agents

Kaibo Huang, Jin Tan, Yukun Wei, Wanling Li, Zipei Zhang, [0.2ex] Hui Tian, Zhongliang Yang, Linna Zhou, [1ex] Beijing University of Posts, Telecommunications · 2026

Read AgentMark: Utility-Preserving Behavioral Watermarking for Agents by Kaibo Huang, Jin Tan, Yukun Wei, Wanling Li, Zipei Zhang, [0.2ex] Hui Tian, Zhongliang Yang, Linna Zhou, [1ex] Beijing University of Posts, Telecommunications (2026) 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-2601-03294.txt.

Description

Kaibo Huang and seven co-authors watermark the behavior of AI agents, not just their words. As agents plan and carry out multi-step tasks on their own, owners need ways to prove provenance and protect ideas. Content watermarking can tag generated text. But it cannot identify the high-level planning choices, such as which tools and subgoals an agent picks, that actually govern execution.

Their framework, AgentMark, embeds multi-bit identifiers into planning decisions while keeping the agent useful. It draws out an explicit behavior distribution from the agent. It then applies distribution-preserving conditional sampling, so decisions carry the mark without drifting in their statistics. The scheme works even through black-box APIs and can combine with ordinary content watermarking. Experiments span embodied, tool-use, and social environments.

Catalog record

CREATORKaibo Huang, Jin Tan, Yukun Wei, Wanling Li, Zipei Zhang, [0.2ex] Hui Tian, Zhongliang Yang, Linna Zhou, [1ex] Beijing University of Posts, Telecommunications
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE73 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-03294

Browse the full library · About the Tower of Babel