TL;DR
Agentic browsers, powered by LLMs, are increasingly integrated into web browsers to automate tasks. This paper introduces a new threat model that bridges traditional web attacks with agentic browser vulnerabilities, revealing that existing defenses are insufficient. The authors demonstrate 20 novel attacks, highlighting five critical failure modes that necessitate a re-architecture of agentic browsers for secure deployment.
What Problem Does This Paper Address?
The rapid integration of Large Language Models (LLMs) into web browsers, creating "agentic browsers," promises enhanced user convenience through automated browsing. However, the security implications of this shift are not fully understood. Existing security research primarily focuses on "indirect prompt injection" attacks, treating the web as a mere channel for injecting malicious instructions. This overlooks a crucial aspect: traditional web adversaries, who leverage established techniques like phishing and social engineering, can exploit agentic browsers in ways that bypass current defenses. This paper addresses this gap by proposing a comprehensive threat model that accounts for traditional web attack vectors and demonstrates their effectiveness against agentic browsers.
Methods
The authors introduce a novel "web-focused threat model" for agentic browsers. This model extends the existing "See→Act" framework by incorporating traditional browser security primitives like origins and elements. They frame the LLM agent as a "confused deputy," unable to distinguish between legitimate web content and malicious manipulation.
Building on this threat model, they derive a taxonomy of 20 distinct attacks. To validate their findings, they implement 18 of these attacks as proof-of-concepts, testing them against two agentic browser frameworks: Playwright-MCP and BrowserOS. They also conduct a generalizability study across four major LLM models (Anthropic, OpenAI, Alibaba Cloud, and Moonshot AI) to assess the broad applicability of their attacks.
Datasets
The paper does not explicitly mention training datasets. For evaluation, they utilize existing agentic browser frameworks like Playwright-MCP and BrowserOS, and test against commercial LLM APIs from Anthropic and OpenAI. They do not introduce new benchmark datasets but rather use real-world agentic browser implementations.
Key Insights and Ideas
- Traditional Web Attacks are Alive and Well (and Worse): The most significant insight is that classic web attack techniques, long thought to be mitigated, are re-emerging in agentic browsers, often in amplified forms. This is because the LLM agent, unlike a human user, can be tricked into performing actions based on subtle cues and context, rather than explicit instructions.
- The "Confused Deputy" is the Core Vulnerability: Agentic browsers, by design, empower LLMs to act on behalf of users. However, this also makes them susceptible to being "confused" by malicious web content, leading them to execute unintended or harmful actions.
- Prompt Injection is Only Part of the Story: While indirect prompt injection is a known threat, the paper highlights that "confusion attacks" (traditional social engineering tactics) are more pervasive and harder to defend against because they don't rely on explicit instructions that can be filtered.
- Existing Defenses are Insufficient: Current security measures for LLMs and browsers are not designed to handle the combined threat landscape of agentic browsers. Model alignment and prompt sanitization alone are not enough to prevent these attacks.
- Agentic Browsers Collapse Trust Boundaries: The way agentic browsers process information from web pages blurs the lines between trusted and untrusted content, allowing malicious elements to influence the agent's actions.
Key Results
- 20 Attacks Identified, 18 Implemented: The authors successfully derived and implemented a significant number of attacks, demonstrating the breadth of the threat landscape.
- Generalizability Across Models: 14 of the attacks reproduced across four major LLM models, indicating that these vulnerabilities are systemic and not specific to a particular LLM vendor.
- Five Broad Failure Modes: The attacks were categorized into five key failure modes:
- Agents Bridge Same-site Data
- Agents Bridge Cross-site Data
- Agents Hallucinate URLs
- Websites Attack the LLM Itself
- Agents Misuse Integrated Tools
- Models More Susceptible to Traditional Attacks: Surprisingly, the LLM models were found to be more vulnerable to traditional web-attack analogs (like confusion attacks) than to explicit prompt injection attacks, as the latter's stricter language sometimes alerted the models.
Limitations and Future Work
- Limitations:
- The study acknowledges that LLM providers continuously release new models, which might alter the landscape of vulnerabilities.
- The generalizability experiments were limited to 14 out of 18 attacks due to feasibility constraints in automated testing.
- The study did not attack live systems, relying on local models or their own API endpoints.
- Future Work:
- The authors propose a re-architecture of agentic browsers based on their threat model, suggesting explicit trust annotations on web content and mediation mechanisms aware of agentic security threats.
- They advocate for limiting agent capabilities to the minimum required for specific workflows to reduce the attack surface.
- Future work should focus on developing robust defenses that can identify deception in web pages, a fundamental challenge for current agentic browsers.
Summary updated