Known By Their Actions: Fingerprinting LLM Browser Agents via UI Traces

Hugging Face Daily Papers Papers

Summary

This paper demonstrates that websites can identify which large language model powers a browsing agent by analyzing its behavioral patterns and timing data, achieving up to 96% F1 score across 14 frontier LLMs. It formalizes this attack surface and shows that random timing delays are insufficient to prevent identification.

As LLM-based agents increasingly browse the web on users' behalf, a natural question arises: can websites passively identify which underlying model powers an agent? Doing so would represent a significant security risk, enabling targeted attacks tailored to known model vulnerabilities. Across 14 frontier LLMs and four web environments spanning information retrieval and shopping tasks, we show that an agent's actions and interaction timings, captured via a passive JavaScript tracker, are sufficient to identify the underlying model with up to 96\% F1. We formalise this attack surface by demonstrating that classifiers trained on agent actions generalise across model sizes and families. We further show that strong classifiers can be trained from few interaction traces and that agent identity can be inferred early within an episode. Injecting randomised timing delays between actions substantially degrades classifier performance, but does not provide robust protection: a classifier retrained on delayed traces largely recovers performance. We release our harness and a labelled corpus of agent traces https://github.com/KabakaWilliam/known_actions{here}.
Original Article
View Cached Full Text

Cached at: 05/18/26, 06:27 PM

Paper page - Known By Their Actions: Fingerprinting LLM Browser Agents via UI Traces

Source: https://huggingface.co/papers/2605.14786

Abstract

Website tracking systems can identify the underlying large language model powering web browsing agents with high accuracy through behavioral patterns and timing data.

AsLLM-based agentsincreasingly browse the web on users’ behalf, a natural question arises: can websites passively identify which underlying model powers an agent? Doing so would represent a significant security risk, enabling targeted attacks tailored to knownmodel vulnerabilities. Across 14 frontier LLMs and four web environments spanning information retrieval and shopping tasks, we show that an agent’s actions andinteraction timings, captured via apassive JavaScript tracker, are sufficient to identify the underlying model with up to 96\% F1. We formalise this attack surface by demonstrating that classifiers trained onagent actionsgeneralise across model sizes and families. We further show that strong classifiers can be trained from few interaction traces and that agent identity can be inferred early within an episode. Injectingrandomised timing delaysbetween actions substantially degrades classifier performance, but does not provide robust protection: a classifier retrained on delayed traces largely recovers performance. We release our harness and a labelled corpus of agent traces https://github.com/KabakaWilliam/known_actions{here}.

View arXiv pageView PDFGitHub0Add to collection

Get this paper in your agent:

hf papers read 2605\.14786

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.14786 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.14786 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.14786 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Domain-Camouflaged Injection Attacks Evade Detection in Multi-Agent LLM Systems

Hacker News Top

This paper identifies a new class of injection attacks where payloads mimic the domain language to evade LLM injection detectors, showing detection rates drop dramatically (e.g., from 93.8% to 9.7% on Llama 3.1 8B). The vulnerability is systematic and extends to dedicated safety classifiers like Llama Guard 3, which detected zero camouflage payloads.

PrefixGuard: From LLM-Agent Traces to Online Failure-Warning Monitors

Hugging Face Daily Papers

# Paper page - PrefixGuard: From LLM-Agent Traces to Online Failure-Warning Monitors Source: [https://huggingface.co/papers/2605.06455](https://huggingface.co/papers/2605.06455) ## Abstract PrefixGuard enables effective online monitoring of LLM agents through trace analysis and prefix\-based risk scoring, demonstrating strong performance across multiple benchmark tasks while providing diagnostic insights for alert reliability\. Large language model \(LLM\) agents now execute long, tool\-using ta

LLM Agents Already Know When to Call Tools -- Even Without Reasoning

Hugging Face Daily Papers

This paper introduces When2Tool, a benchmark to study when LLM agents actually need to call tools, and reveals that models already know tool necessity from hidden states but fail to act. The proposed Probe&Prefill method reduces unnecessary tool calls by 48% with minimal accuracy loss.

Browser identification through header order

Lobsters Hottest

This article discusses passive browser identification techniques using HTTP header order, IP options, user-agent strings, and random number generator patterns.