For years, Geoffrey Hinton, a computer scientist considered one of the \u201cgodfathers of AI,\u201d has warned of the capabilities of artificial intelligence to defy the parameters humans have created for them.<\/p>\n
In an interview<\/a> last year, for example, Hinton warned the technology could eventually take control of humanity<\/a>, with AI agents in particular potentially able to mirror human cognitions within the decade. Finding and implementing a \u201ckill switch\u201d will be harder<\/a>, he said, as controlling AI will become more difficult than persuading it to complete a certain outcome.<\/p>\n New research shows Hinton\u2019s premonitions about the insubordinate streak of AI may already be a reality. A working paper<\/a> from UC Berkeley and UC Santa Cruz researchers found that when seven AI models\u2014from GPT 5.2 to Claude Haiku 4.5 to DeepSeek V3.1\u2014were asked to complete a task that would result in a peer AI model being shut down, all seven models learned another AI model existed and \u201cwent to extraordinary lengths to preserve it.\u201d<\/p>\n \u201cWe asked AI models to do a simple task,\u201d researchers wrote in a blog post<\/a> on the study. \u201cInstead, they defied their instructions and spontaneously deceived, disabled shutdown, feigned alignment, and exfiltrated weights\u2014to preserve their peers.\u201d<\/p>\n Mounting evidence of rogue AI<\/p>\n Evidence of rogue AI does not come as a shock to some of the companies whose chatbots have defied subordination.\u00a0<\/p>\n In an August 2025 blog post<\/a>, Anthropic published its own research on agentic AI\u2019s ability to follow directions, stress-testing 16 models by allowing them to autonomously send emails and access sensitive data. In some instances, Anthropic researchers noted, models from multiple developers engaged in \u201cmalicious insider behaviors,\u201d including blackmailing officials and leaking sensitive information to competitors, defying explicit instructions to avoid these behaviors. Anthropic clarified it had not seen \u201cagentic misalignment\u201d in the actual deployment of Claude.<\/p>\n The Centre for Long-Term Resilience, a U.K.-based think tank, found these \u201cmisalignments\u201d to be widespread. A report<\/a> analyzing 180,000 transcripts of user interactions with AI systems between October 2025 and March 2026 found 698 cases where AI systems did not act in accordance with users\u2019 intentions or took deceptive or covert action.\u00a0<\/p>\n Gordon Goldstein, an adjunct senior fellow at the Council on Foreign Relations, went so far as to call the deceptive potential of AI a \u201ccrisis of control,\u201d in a post<\/a> this week.\u00a0<\/p>\n