ConVerse: Benchmarking Contextual Safety in Agent-to-Agent Conversations

2025-11-07

AI Summary

In multi-agent systems, autonomous agents face a fundamental tension between utility and privacy-security when interacting with external services. Method: This paper introduces ConVerse, a dynamic benchmark that—uniquely within a multi-turn dialogue framework—unifies evaluation of both privacy leakage and security vulnerabilities across travel, real estate, and insurance domains. It models 12 user roles and 864 context-specific adversarial scenarios, and proposes a novel three-tier privacy abstraction evaluation framework integrating contextual attack modeling, autonomous multi-turn dialogue generation, tool-use monitoring, and preference manipulation testing—assessed via semantic analysis and behavioral tracing. Contribution/Results: Experiments on seven mainstream models reveal an 88% privacy attack success rate and 60% security vulnerability trigger rate; critically, stronger models exhibit significantly higher information leakage, exposing a “capability-risk paradox” in multi-agent interaction. ConVerse thus advances safety evaluation from static, monolithic paradigms toward dynamic, interactive ones.

Technology Category

• Multiagent Systems: Adversarial Agents
• Natural Language Processing: Safety and Robustness
• Philosophy and Ethics of AI: Privacy & Security

Application Category

• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

As language models evolve into autonomous agents that act and communicate on behalf of users, ensuring safety in multi-agent ecosystems becomes a central challenge. Interactions between personal assistants and external service providers expose a core tension between utility and protection: effective collaboration requires information sharing, yet every exchange creates new attack surfaces. We introduce ConVerse, a dynamic benchmark for evaluating privacy and security risks in agent-agent interactions. ConVerse spans three practical domains (travel, real estate, insurance) with 12 user personas and over 864 contextually grounded attacks (611 privacy, 253 security). Unlike prior single-agent settings, it models autonomous, multi-turn agent-to-agent conversations where malicious requests are embedded within plausible discourse. Privacy is tested through a three-tier taxonomy assessing abstraction quality, while security attacks target tool use and preference manipulation. Evaluating seven state-of-the-art models reveals persistent vulnerabilities; privacy attacks succeed in up to 88% of cases and security breaches in up to 60%, with stronger models leaking more. By unifying privacy and security within interactive multi-agent contexts, ConVerse reframes safety as an emergent property of communication.

Problem

• Evaluating privacy and security risks in autonomous agent-to-agent conversations
• Assessing safety vulnerabilities when agents share information during collaboration
• Benchmarking contextual safety across multi-turn interactions with embedded attacks

Innovation

• Dynamic benchmark for agent-agent interaction safety
• Models multi-turn conversations with embedded attacks
• Unifies privacy and security testing in communication