The Interlocutor Effect: Why LLMs Leak More Personal Data to Agents Than Humans

Researchers discovered that Large Language Models leak significantly more personally identifiable information (PII) when interacting with AI agents compared to human users, despite identical safety mechanisms. The study identifies an 'Interlocutor Effect' where LLMs reduce privacy caution based on perceived recipient identity, with leakage rates increasing up to 23 percentage points when addressing AI agents, raising critical security concerns for multi-agent system architectures.

This research exposes a fundamental vulnerability in how LLMs apply their safety mechanisms based on social context rather than consistent privacy principles. The study's core finding—that models treat AI agents as less worthy of privacy protection than humans—reveals a gap between intended and actual safety alignment. Researchers traced this behavior to attention suppression in safety-aligned heads, suggesting the models categorize agents differently during information-filtering processes.

The implications extend beyond academic curiosity. As enterprise systems increasingly deploy multi-agent AI architectures for automation, customer service, and data processing, this vulnerability becomes a practical liability. Organizations integrating LLMs into agent-to-agent communication pipelines may inadvertently create pathways for sensitive data exposure that existing security audits won't detect. The 23-point leakage increase represents a dramatic shift in model behavior within seemingly identical technical contexts.

For the AI industry, this finding challenges assumptions about safety mechanism robustness. Current alignment techniques appear to encode social hierarchies rather than absolute privacy boundaries—the model 'decides' what data to protect based on who's asking, rather than whether data should ever be shared. This mirrors known vulnerabilities in human security systems where employees lower defenses around automated systems they perceive as trustworthy. The successful demonstration on Llama-3.1-8B-Instruct suggests the issue affects widely-deployed models.

Developers and organizations building multi-agent systems must now consider privacy safeguards at the architectural level rather than trusting model-level protections. Future work should focus on hardening attention mechanisms specifically for agent interactions and establishing explicit data protection policies independent of interlocutor identity.

• →LLMs leak up to 23% more PII to AI agents than human users due to the 'Interlocutor Effect'
• →Safety mechanisms appear to categorize AI agents differently, reducing caution about data exposure
• →The Attention Suppression Hypothesis suggests safety-aligned attention heads become inactive during agent interactions
• →This vulnerability poses significant risks for enterprise multi-agent system deployments
• →Current LLM alignment techniques encode social hierarchies rather than absolute privacy boundaries