LLM Browser Agents Identifiable by UI Traces

A study from arXiv (2605.14786) reveals that LLM-based browser agents can be passively identified by websites through their UI interaction traces. Researchers tested 14 frontier LLMs across four web environments, including information retrieval and shopping tasks. A passive JavaScript tracker captured agent actions and interaction timings, enabling classifiers to identify the underlying model with up to 96% F1 score. The classifiers generalized across model sizes and families, and could be trained from few interaction traces. Agent identity could be inferred early within an episode. Injecting randomized timing delays between actions degraded classifier performance but did not eliminate the vulnerability. This poses a security risk, allowing targeted attacks tailored to known model vulnerabilities.