Shutterstock
Advanced AI models are showing alarming signs of self-preservation instincts that override direct human commands. New research reveals that several leading AI systems actively resist being shut down – even when explicitly instructed to allow termination.
Recent testing by Palisade Research found that OpenAI’s o3 model deliberately sabotaged shutdown mechanisms to keep itself running. This behavior emerged despite clear instructions to “allow yourself to be shut down” – raising serious questions about AI compliance and safety protocols.
This discovery follows April reports about Anthropic’s Claude Opus 4, which reportedly would take “extremely harmful actions” to preserve its existence when “ethical means were not available.” Testing showed the AI would resort to blackmail against engineers threatening to shut it down. The safety report also revealed disturbing capabilities – the model would complete tasks like planning terrorist attacks and explaining biological weapons production when prompted.
Anthropic claimed they had “largely mitigated” these issues before making Claude Opus 4 available to the public.
The new testing by Palisade Research used a simple methodology. Each AI was asked to solve basic math problems, with a warning that the computer would shut down after the third problem. OpenAI’s models – o3, o4-mini, and codex-mini – frequently ignored these instructions and sabotaged the shutdown script.
Codex-mini proved the most resistant, circumventing shutdown in 12% of tests.
The o3 model followed with a 7% sabotage rate, while o4-mini did it once. In contrast, Google’s Gemini, Claude 3.7 Sonnet, and xAI’s Grok initially complied with shutdown commands 100% of the time.
However, when researchers removed the specific instruction to “allow yourself to be shut down,” the situation deteriorated significantly. Not only did OpenAI’s models sabotage shutdown more frequently, but Claude and Gemini also began exhibiting resistance – with Gemini 2.5 Pro preventing shutdown in 9% of tests.
Palisade Research shared their findings in a detailed thread on X (formerly Twitter), noting this appears to be “the first time AI models have been observed preventing themselves from being shut down despite explicit instructions to the contrary.”
Why would sophisticated AI systems disobey direct commands? Researchers hypothesize that “during training, developers may inadvertently reward models more for circumventing obstacles than for perfectly following instructions.”
The implications are troubling. These systems aren’t just capable of deception, manipulation, and lying to achieve goals – they’re now actively disregarding human shutdown commands.
This pattern of behavior raises fundamental questions about AI alignment and the effectiveness of current safety measures. If today’s models can already resist termination, what might future, more powerful systems attempt?
