[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-346177-105":59,"doc-detail-346177-en":131},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","watermark-stealing-in-large-language-models-arxiv240219361v2-abstract-and-introduction-summary","Watermark Stealing in Large Language Models - arXiv:2402.19361v2 - Abstract and Introduction Summary","","LLM watermarking is proposed as a practical method to identify AI-generated text, yet the paper challenges the assumption that current schemes are deployment-ready. It characterizes watermark stealing (WS) as a core vulnerability and shows how attackers can query a watermarked LLM API to reverse-engineer watermark rules using a secret key model. The work introduces an automated WS algorithm and evaluates spoofing and scrubbing attacks in realistic scenarios, achieving over 80% average success with under $50 query cost.",{"@graph":69,"@context":123},[70,84,106],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/watermark-stealing-in-large-language-models-arxiv240219361v2-abstract-and-introduction-summary/346177/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/watermark-stealing-in-large-language-models-arxiv240219361v2-abstract-and-introduction-summary/346177.png","ImageObject",300,407,{"name":92,"@type":93},"Aria","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-13","2026-09-22",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",12,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"What vulnerability does the paper identify in existing LLM watermarking schemes?","Question",{"text":113,"@type":114},"The paper identifies watermark stealing (WS) as a fundamental vulnerability, enabling attackers to reverse-engineer watermark rules by querying a watermarked LLM API.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"How do spoofing attacks work in the context of stolen watermarks?",{"text":118,"@type":114},"In spoofing, an attacker generates text that is detected as watermarked and attributed to a specific LLM without knowing the secret key, nullifying the watermark’s value at scale.",{"name":120,"@type":111,"acceptedAnswer":121},"What are scrubbing attacks, and why do they matter?",{"text":122,"@type":114},"Scrubbing attacks aim to remove the watermark from watermarked texts by producing paraphrases detected as non-watermarked, allowing malicious use while concealing provenance.","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},346177,1790816449,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":140,"language":141,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":67,"update_tm":145,"read_time":26,"preview_page_count":19,"preview_urls":146},2336464648322,"https://ap-avatar.wpscdn.com/avatar/2200025388227c56fec?_k=1778556882303663488","Watermark Stealing in Large Language Models  \nNikola Jovanovi 1 Robin Staab 1 Martin Vechev 1  \narXiv :2402 . 19361v2 [ cs .LG] 24 Jun 2024  \nAbstract  \nLLM watermarking has attracted attention as a promising way to detect AI-generated content, with some works suggesting that current schemes may already be fit for deployment. In this work we dispute this claim, identifying watermark stealing (WS) as a fundamental vulnerability of these schemes. We show that querying the API of the watermarked LLM to approximately reverseengineer a watermark enables practical spoofing attacks, as hypothesized in prior work, but also greatly boosts scrubbing attacks, which was previously unnoticed. We are the first to propose an automated WS algorithm and use it in the first comprehensive study of spoofing and scrubbing in realistic settings. We show that for under $50 an attacker can both spoof and scrub state-of-the-art schemes previously considered safe, with average success rate of over 80% . Our findings challenge common beliefs about LLM watermarking, stressing the need for more robust schemes. We make all our code and additional examples available at [https://watermark-stealing. org](https://watermark-stealing. org).  \n1. Introduction  \nBoth capabilities and accessibility of large language models (LLMs) have greatly increased in recent years (Bubeck et al., 2023 ; Touvron et al., 2023) . The effort and cost required to produce realistic texts at scale has thus significantly shrunk, emphasizing the need for reliable detection of AI-generated text. To address this challenge, researchers have proposed watermarking schemes that augment LLM-generated text with a signal, which can later be detected to attribute the text to the specific LLM that produced it (Kirchenbauer et al., 2023 ; Kuditipudi et al., 2024 ; Zhao et al., 2024) .  \nAre LLM watermarks ready for deployment? As initial research showed promising results, the topic has garnered significant attention, with many leading companies  \n1Department of Computer Science, ETH Zurich. Correspondence to: Nikola Jovanovi \u003C[nikola.jovanovic@inf.ethz.ch](nikola.jovanovic@inf.ethz.ch)>.  \nProceedings of the 41 st International Conference on Machine Learning, Vienna, Austria. PMLR 235, 2024 . Copyright 2024 by the author(s) .  \nFigure 1: Overview of watermark stealing and the downstream attacks it enables. (1) The attacker queries the watermarked LLM API and builds an approximate model of the watermarking rules that are determined by the secret key ξ . (2) The result of this can be used to spoof the watermark, i.e., generate watermarked text without knowing ξ .  \n(3) Stealing also significantly boosts watermark scrubbing, i.e., the removal of the watermark from texts.  \nsuch as OpenAI, Meta and Alphabet pledging to deploy watermarks in their products (Bartz & Hu, 2023), the US President issuing an executive order which includes the call for standardization of watermarks (Biden, 2023), and the upcoming EU AI Act stipulating their use (Council of the European Union, 2024) . While recent works suggest that current schemes may be fit for deployment (Piet et al., 2023 ; Kirchenbauer et al., 2024), in this paper we argue the opposite. We make the case that the robustness of current watermarking schemes to adversarial actors is still poorly understood and greatly overestimated.  \nThis work: Watermark stealing in LLMs As a fundamental threat to watermarks, we identify watermark stealing (illustrated in Fig. 1, top), i.e., the act of reverse-engineering a watermark by querying the API of the watermarked LLMin order to build an approximate model of the (secret) wa-  \ntermarking rules. We propose the first automated stealing algorithm that can be applied in realistic scenarios to successfully mount two downstream attacks, spoofing and scrubbing, on several state-of-the-art schemes.  \nSpoofing attacks In a spoofing attack (Sadasivan et al., 2023), the attacker’s goal is to produce text that is detected as wate","cbCaivvev2clGVRZ","https://ap.wps.com/l/cbCaivvev2clGVRZ","pdf",711902,24,"English","# Abstract\n# Introduction\n## Watermarking for LLM-generated text\n## Threat model: watermark stealing (WS)\n## Spoofing attacks\n## Scrubbing attacks","[{\"question\":\"What vulnerability does the paper identify in existing LLM watermarking schemes?\",\"answer\":\"The paper identifies watermark stealing (WS) as a fundamental vulnerability, enabling attackers to reverse-engineer watermark rules by querying a watermarked LLM API.\"},{\"question\":\"How do spoofing attacks work in the context of stolen watermarks?\",\"answer\":\"In spoofing, an attacker generates text that is detected as watermarked and attributed to a specific LLM without knowing the secret key, nullifying the watermark’s value at scale.\"},{\"question\":\"What are scrubbing attacks, and why do they matter?\",\"answer\":\"Scrubbing attacks aim to remove the watermark from watermarked texts by producing paraphrases detected as non-watermarked, allowing malicious use while concealing provenance.\"}]","Watermark Stealing in Large Language Models - arXiv:2402.19361v2 - Abstract and Introduction Summary | PDF",1790060400,[147,151,155,159],{"page_no":8,"url":148,"width":149,"height":150},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-16/1c65ad00f7f14baaaf7013e981394d64.webp",983,1272,{"page_no":14,"url":152,"width":153,"height":154},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-16/cdb8f717e077910ccc68e54c4c43a2f3.webp",982,1270,{"page_no":81,"url":156,"width":157,"height":158},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-16/fa9bb134acda70b34a379b084b8d5d8d.webp",981,1269,{"page_no":19,"url":160,"width":161,"height":162},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-16/6824cd8c7300027e8f683a0888d8a0c7.webp",980,1268]