AgentDAM: Privacy Leakage Evaluation for Autonomous Web Agents

Arman Zharmagambetov, Chuan Guo, Ivan Evtimov, Maya Pavlova, Ruslan Salakhutdinov, Kamalika Chaudhuri

🏛 Institutions: FAIR at Meta, Meta
📅 Date: March 12, 2025
📑 Publisher: NeurIPS 2025 Datasets and Benchmarks Track (Poster)
💻 Env: Web
🔑 Keywords: benchmark privacy data minimization prompting-based defense AgentDAM

TLDR

AgentDAM is an end-to-end web benchmark for testing whether autonomous agents obey data minimization and avoid accessing sensitive information unless it is necessary for the task. It shows that current agents frequently over-consume private data, while a prompting-based defense can reduce leakage.

Open paper arXiv Edit on GitHub Report issue