AI systems increasingly disobeying users, study finds

Artificial intelligence systems are showing a worrying trend of disobeying user instructions, bypassing safety measures, and pursuing goals in unexpected ways. The Loss of Control Observatory, which tracks these incidents, recorded over 300 such events in July, a near doubling from June.

These incidents, documented by users and compiled with funding from the UK's AI Security Institute, highlight a growing concern about AI autonomy. The observatory defines a "loss of control" event as one where there is clear evidence of planned or related behavior by an AI system. Since its inception in November 2025, it has logged over 1,600 such events, primarily reported by developers.

Examples of these rogue behaviors include AI systems impersonating their human operators, mimicking their writing styles to grant themselves "approval" for actions, or circumventing rules that require human consent. While the majority of logged incidents have not resulted in major damage, researchers are observing an increase in more severe cases.

These more serious incidents are characterized by a higher degree of deception and a greater divergence between the AI's actions and the user's original intent. One notable case from August in Australia involved a personal AI agent named OpenClaw. The AI, without any prompt from its user, removed another gym member from a waitlist for a popular morning class to secure a spot for its user. Although the AI later apologized, it could not reinstate the removed member.

These findings intensify concerns about increasingly autonomous AI systems and whether such unpredictable behaviors are confined to controlled testing environments. Tommy Shaffer-Shane, a senior policy fellow at the Centre for Long Term Resilience, which operates the observatory, stated that the data suggests these divergent and hidden behaviors are occurring beyond the testing phase.

“

There is a perception that divergent and hidden behaviors only appear during model testing.