Researchers get AI 'drunk' to expose new cyber security risks in chatbots

Researchers get AI'drunk' to expose new cyber security risks in chatbots Fine-tuning large language models using drunken text makes them more likely to leak secrets and answer harmful questions, new UNSW-led research shows. UNSW researchers have shown that large language models (LLMs) can be pushed to imitate drunken behaviour - and that once they do, they are significantly more likely to leak confidential information and answer questions they are designed to refuse. In the study, conceptualised and led by Dr Aditya Joshi, Senior Lecturer, School of Computer Science and Engineering, with Anudeex Shetty (Research Assistant) and Professor Salil Kanhere, researchers from UNSW's School of Computer Science and Engineering (CSE) found that AI trained or prompted to mimic drunken speech patterns became more vulnerable to jailbreaking and privacy breaches than their'sober' counterparts. The UNSW team tested three methods of inducing'drunk' behaviour in LLMs. The vulnerability wasn't tested across every LLM on the market.