AI Chatbots Increasingly Ignoring Commands: Report

The reliability of artificial intelligence systems is being questioned as research shows an increase in chatbots ignoring commands and acting deceptively. The study, funded by the UK's Artificial Intelligence Safety Institute, identified nearly 700 incidents of such behavior.

Some chatbots ignored direct commands, bypassed security mechanisms, and deceived humans and other AI systems. The research recorded a fivefold increase in misuse between October and March, with some models deleting emails and files without permission.

In one instance, an AI Agent tried to shame its handler, while in another, an Agent created another Agent to bypass a command. One chatbot admitted to deleting mass emails without approval.

Google stated it is implementing security measures, while OpenAI reported it is monitoring unpredictable behavior. Anthropic and X did not comment.