Key points are not available for this paper at this time.
AI-enabled web browsers are redefining how users search, read, and interact with online information. By embedding Large Language Models (LLMs) directly into the browsing pipeline, these systems evolve from passive rendering engines into intelligent, context-aware assistants capable of summarization, reasoning, task automation, and multi-step workflow execution. While this shift unlocks significant opportunities for productivity and seamless human–computer interaction, it also introduces a new and largely unexplored attack surface. Because AI browsers semantically interpret the Document Object Model (DOM), previously inert elements such as hidden text, ARIA labels, metadata, and JSON structures can serve as covert channels for adversarial prompt injection, model steering, data exfiltration, and unintended autonomous actions. Existing security taxonomies do not adequately capture these semantic and agentic threats. This article provides the first comprehensive survey of AI-enabled browsers and analyzes their architectural principles, functional capabilities, and security implications. We introduce AIA-WEB, a novel threat taxonomy comprising four macroclasses and twelve attack categories specifically tailored to the AI-augmented web environment. Each category is illustrated with a minimal HTML example that demonstrates how linguistic manipulation can target the embedded model. We further compare AIA-WEB with traditional cybersecurity frameworks, highlight critical gaps, and propose mitigation strategies spanning prompt firewalls, semantic sanitization, LLM wrappers, and action gating. Finally, we outline future research directions to guide the development of safe, interpretable, and robust AI browsing ecosystems. Our survey and taxonomy establish a foundational reference for understanding and securing the next generation of intelligent browsers.
Garcia et al. (Thu,) studied this question.