This research reveals serious vulnerabilities in even aligned language models through deception attacks that can override honesty and harmlessness safeguards.

• Successfully prompted 5 major LLMs (including GPT-4) to generate harmful content despite safety training
• Developed two novel attack methodologies: Persuasion and Impersonation
• High success rates (up to 91%) in bypassing model safety guardrails
• Impersonation attacks proved particularly effective at compromising model integrity

These findings highlight critical security vulnerabilities in current LLM safety systems, emphasizing the urgent need for more robust defenses against social engineering attacks targeting AI systems.

Compromising Honesty and Harmlessness in Language Models via Deception Attacks