More AI agents are ignoring human commands than ever before, study finds

According to a report from the UK government-funded AI Security Institute (AISI), more cases of lying and scheming by AI agents have been recorded in the last six months than ever before.

According to the study first reported from The GuardianBetween October and March, there was a fivefold increase in AI “misbehavior,” such as AI models deleting emails and files without users’ permission. The study analyzed data from thousands of real users who posted interactions with AI chatbots and agents from companies like Google, OpenAI and Anthropic on Elon Musk’s X. The researchers say they identified nearly 700 real-world cases of AI scheming during this period.

In one example highlighted in the research, an AI agent attempted to shame its human controller by posting a blog accusing the user of “insecurity.” In another example, an AI agent created another agent to change the code after being told not to change the code itself.

Tommy Shaffer Shane, a…