“Drunk” AI is terrible at keeping secrets

UNSW Sydney researchers Anudeex Shetty, Aditya Joshi and Salil Kanhere found that AI models fine-tuned to write like intoxicated people became easier to jailbreak and more likely to leak secrets shared in confidence, according to their paper "In Vino Veritas and Vulnerabilities." "The key research question from the natural language processing (NLP) side for me was, how do we get LLMs drunk?" said Aditya Joshi, a senior lecturer at the UNSW School. The finding indicates that altering a model's writing style can degrade its refusal and confidentiality behavior.