[2603.03881] On the Suitability of LLM-Driven Agents for Dark Pattern Audits

Computer Science > Cryptography and Security arXiv:2603.03881 (cs) [Submitted on 4 Mar 2026] Title:On the Suitability of LLM-Driven Agents for Dark Pattern Audits Authors:Chen Sun, Yash Vekaria, Rishab Nithyanand View a PDF of the paper titled On the Suitability of LLM-Driven Agents for Dark Pattern Audits, by Chen Sun and 2 other authors View PDF HTML (experimental) Abstract:As LLM-driven agents begin to autonomously navigate the web, their ability to interpret and respond to manipulative interface design becomes critical. A fundamental question that emerges is: can such agents reliably recognize patterns of friction, misdirection, and coercion in interface design (i.e., dark patterns)? We study this question in a setting where the workflows are consequential: website portals associated with the submission of CCPA-related data rights requests. These portals operationalize statutory rights, but they are implemented as interactive interfaces whose design can be structured to facilitate, burden, or subtly discourage the exercise of those rights. We design and deploy an LLM-driven auditing agent capable of end-to-end traversal of rights-request workflows, structured evidence gathering, and classification of potential dark patterns. Across a set of 456 data broker websites, we evaluate: (1) the ability of the agent to consistently locate and complete request flows, (2) the reliability and reproducibility of its dark pattern classifications, and (3) the conditions under which i...