Researchers have recently discovered a critical flaw in large language models, the ones powering chatbots like ChatGPT and Claude: it takes far less work to sabotage these AIs than experts had previously believed.
