Browser agents automate complex workflows by reading and acting on arbitrary page content, but that flexibility exposes them to prompt injection attacks and unintended data exfiltration. A new arXiv preprint, SPIDES, takes an information-flow control approach to this problem, treating it as a data-flow issue rather than just an LLM robustness issue.
Information-flow control is a classic security technique that tracks how data moves through a system and restricts where sensitive data can go. Applied to browser agents, it could prevent a malicious webpage from steering the agent into leaking data or performing unintended actions. The abstract also references recent work on LLM-driven browser agents, though that sentence is cut off in the announcement.
The paper is a preprint, so it has not yet been peer-reviewed, and the abstract alone does not provide implementation details or evaluation results. Still, the direction is notable: instead of trying to make the LLM itself more resistant to injection, SPIDES aims to enforce boundaries on how data flows through the agent's actions.