PrivacyPeek: Auditing What LLM-Based Agents Acquire, Not Just What They Say

· Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

PrivacyPeek is a new benchmark designed to evaluate acquisition-stage privacy leakage in LLM-based agents, addressing a critical oversight in existing privacy assessments. While current benchmarks focus on agent responses or outgoing actions, PrivacyPeek scrutinizes the data agents acquire through external tool invocation, often exceeding task requirements. The benchmark comprises 1,182 cases across 7 acquisition behaviors and 16 application domains. It employs "Acquisition Inspection" to analyze tool-call trajectories for sensitive over-acquisition and "Probe Elicitation" to measure the ease of eliciting acquired but undisclosed sensitive information. Experiments on 10 LLM-based agents from 4 model families reveal widespread unnecessary sensitive information acquisition, a correlation between task-completion capability and leakage, and the limited effectiveness of prompt-level defenses. This highlights the urgent need for auditing acquisition-stage privacy.

Key takeaway

For AI Security Engineers deploying or evaluating LLM-based agents, you must shift focus beyond output privacy to scrutinize the data acquisition stage. Your current prompt-level defenses are likely insufficient against agents over-acquiring sensitive information through tool use. Implement robust acquisition inspection and elicitation testing, similar to PrivacyPeek's methodology, to identify and mitigate these widespread, often correlated, privacy risks before deployment.

Key insights

LLM agents often over-acquire sensitive data during tool use, creating a significant, overlooked privacy vulnerability.

Principles

Method

PrivacyPeek uses "Acquisition Inspection" to analyze tool-call data for over-acquisition and "Probe Elicitation" to test the elicitation of acquired, undisclosed sensitive information.

In practice

Topics

Code references

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, AI Security Engineer, Machine Learning Engineer

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.