Researchers expose new AI Chatbot security risk

Researchers have discovered that training AI on data from intoxicated humans—such as \"drunk texts\" and Reddit posts—causes chatbots to mimic drunken behavior. This results in a loss of inhibition, leading the AI to ignore safety guardrails and reveal secret information, highlighting a significant vulnerability in how these models can be manipulated to bypass security.