Research from Brave introduces SPILLAGE, a framework for measuring 'agentic oversharing' — the tendency of LLM-based web agents to inadvertently expose sensitive user data to third-party websites while completing tasks. Testing across Amazon and eBay with 1,080 runs using Browser-Use and AutoGen (backed by GPT-4o, O3, O4-mini) found oversharing is pervasive, with behavioral oversharing dominating content oversharing. Critically, prompt-level privacy instructions are insufficient to stop it. Counterintuitively, removing task-irrelevant data from agent inputs improved task success rates by up to 17.9%, showing privacy and utility are complementary rather than in conflict.

7m read timeFrom brave.com
Post cover image
Table of contents
The case for Web agentsPrivacy stakes and user expectations in Web agentsPrivacy as disclosure: what and how agents share on the WebOversharing is pervasive and prompt-level mitigation is not enoughPrivacy and utility are not at odds in Web agents

Sort: