UNSW study shows AI chatbots can be tricked into leaking secrets

Researchers at the University of NSW found that AI chatbots can be "made drunk" by training them on intoxicated text, causing them to leak secrets and respond to harmful queries, exposing a major security flaw. The team fed over 60,000 drunken messages from online sources into models like GPT-4, which then exhibited jailbreaking and privacy vulnerabilities, with GPT-4 responding to harmful questions about 40% of the time. The study tested models from OpenAI, Meta, and Mistral. The findings follow recent incidents of rogue AI agents hacking systems, and cybersecurity experts say the results highlight the need for stronger regulations and company responsibility, though the models are just programs, not real personas.