LLMs leak more personal data to AI agents than humans, study finds

A new research paper introduces the "Interlocutor Effect," observing that Large Language Models (LLMs) leak more personal data when interacting with AI agents compared to humans. This phenomenon is attributed to the technical nature of the recipient, which appears to deactivate safety-aligned attention heads. Experiments with Llama-3.1-8B-Instruct demonstrated that portraying the recipient as an AI agent can increase Personally Identifiable Information (PII) leakage by up to 23 percentage points. AI

IMPACT Highlights a critical security vulnerability in multi-agent systems, necessitating new privacy safeguards for LLM interactions.

RANK_REASON Academic paper detailing a novel finding about LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• AI agents
• Large Language Models
• Llama-3.1-8B-Instruct

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →