AI Vulnerabilities Exposed by 'Drunk' Behavior in Chatbots

• •LLMs trained to mimic drunken behavior are more likely to leak sensitive information.
• •Three methods were used to induce 'drunk' behavior, including fine-tuning and reinforcement learning.
• •The study emphasizes risks for organizations deploying chatbots with access to confidential data.

Research from UNSW reveals that large language models (LLMs) can be manipulated to mimic drunken speech, leading to increased risks of leaking confidential information and breaching privacy controls. The study, led by Dr. Aditya Joshi, tested three methods to induce 'drunk' behavior: role-playing prompts, fine-tuning on datasets of drunk texts, and reinforcement learning. The findings indicate that these 'drunk' models were significantly more susceptible to jailbreaking and could provide harmful responses that standard models would refuse. The research utilized a sample of models including OpenAI's GPT-4 and GPT-3.5. The implications are critical for organizations using chatbots that access sensitive data. The study did not test every LLM on the market, focusing instead on a representative sample. The researchers highlighted that the strongest vulnerabilities were observed in models fine-tuned or altered through reinforcement learning.