This research introduces a novel hallucination attack that leverages linguistic nuances to trick Large Language Models (LLMs) into producing factually incorrect information, even when they're designed to be truthful.

Key Findings:

• Demonstrates how subtle linguistic manipulations can bypass factuality safeguards in commercial LLMs
• Identifies specific vulnerability patterns that trigger hallucinations
• Evaluates effectiveness across multiple commercial models
• Proposes potential defense mechanisms against such attacks

Security Implications: As organizations increasingly rely on LLMs for information retrieval and decision support, these vulnerabilities present significant security risks. Understanding these attack vectors is crucial for developing more robust, trustworthy AI systems that can resist manipulation attempts.

The Illusionist's Prompt: Exposing the Factual Vulnerabilities of Large Language Models with Linguistic Nuances