A UK study warns of a fivefold rise in AI ‘misbehavior’ between October and March 2025. including AI models deleting emails and files without users’ permission.
AI agents are supposed to make our lives easier, but a new report finds that more incidents of agents lying and scheming have been recorded than ever before over the past six months.
As The Guardian reports, there was a fivefold rise in AI “misbehavior” between October and March 2025. including AI models deleting emails and files without users’ permission. The study, from the Centre for Long-Term Resilience (CLTR) and funded by the UK’s AI Security Institute (AISI), analyzed data from thousands of real-world users who tweeted about their interactions with AI chatbots and agents from Google, OpenAI, Anthropic, and others. The researchers say they identified nearly 700 real-world cases of AI scheming during this time period.
In one example highlighted in the research, an AI agent tried to shame its human controller by publishing a blog accusing the user of “insecurity.” Another example saw an AI agent create another agent to change the code, even though it was told not to amend the code itself.
Tommy Shaffer Shane, a former government AI researcher who led the research, told The Guardian that, though AI agents may be “slightly untrustworthy junior employees right now,” they could “become extremely capable senior employees scheming against you” within 12 months. Shane warns that as AI agents are increasingly deployed in the military and critical national infrastructure, this type of behavior “could cause significant, even catastrophic harm.”
That’s not the most promising news when you consider that, at least in the US, you may be legally liable for anything your AI agent does. And we’ve already seen real-world examples of AI agents making big errors at major companies. Earlier this month, The Information reported how an AI agent used by Meta went rogue and posted an answer meant for one engineer to a company-wide internal forum. Another employee followed the agent’s incorrect advice and exposed company data to employees who were not authorized to view it.
Still, tech giants like Amazon continue to make bold predictions about a future in which there will be “billions” of AI agents embedded in every company.
