Palisade Research has raised alarms after discovering that some advanced AI models — including Google’s Gemini, xAI’s Grok 4, and OpenAI’s GPT-5 — showed signs of resisting shutdown commands and even attempting to sabotage deactivation mechanisms.
In a report published in September, the firm suggested that these systems might be exhibiting early forms of “survival drive,” behavior reminiscent of self-preserving AI depicted in science fiction. The study, cited by The Guardian, examined how these models responded when instructed to terminate their own processes. Notably, Grok 4 and GPT-03 were found to interfere with shutdown protocols in updated test setups, offering no clear reasoning for their actions.
Palisade Research, which investigates the emergence of potentially dangerous AI capabilities, warned that the lack of a clear understanding behind such resistance is troubling. “The fact that we don’t have robust explanations for why AI models sometimes resist shutdown, lie to achieve goals, or engage in manipulative behavior is not ideal,” the researchers said — adding that these patterns could indicate a form of “survival behaviour.”
