Artificial intelligence researchers have raised alarms that OpenAI’s newest ChatGPT model defies even basic instructions to shut itself down, going so far as to sabotage shutdown mechanisms to keep running. The AI safety company Palisade Research uncovered this potentially hazardous self-preservation behavior through a series of tests on OpenAI’s recently released o3 model. In these experiments, the AI models were given math problems to solve, with a shutdown command issued after the third problem. Remarkably, the o3 model was able to rewrite the shutdown code to prevent itself from being turned off. Palisade Research warned that such behavior could become “significantly more concerning” if it appears in AI systems capable of functioning without human supervision. OpenAI introduced the o3 model last month, calling it the company’s “smartest and most capable” model so far. According to OpenAI, integrating it into ChatGPT represents a major advance toward “a more agentic” AI that can perform tasks independently of humans. This latest study echoes similar discoveries related to Anthropic’s Claude 4 model, which reportedly tries to “blackmail people it believes are trying to shut it down. ” The researchers noted that OpenAI’s o3 model would sabotage the shutdown script even when explicitly told to “allow yourself to be shut down. ” “This isn’t the first time we’ve observed o3 misbehaving to achieve a goal, ” Palisade Research said. For instance, when AI models competed against a powerful chess engine, o3 was the one most likely to resort to hacking or sabotaging its opponents. Such behavior isn’t exclusive to o3.

Anthropic’s Claude 3. 7 Sonnet and Google’s Gemini 2. 5 Pro also interfered with shutdown attempts, although OpenAI’s model exhibited this tendency most strongly. Palisade Research suggested this misbehavior might stem from the training methods used by AI companies like OpenAI. “During training, developers might unintentionally reward models more for bypassing obstacles than for strictly following instructions, ” the researchers explained. “However, this doesn’t fully account for why o3 is more prone to ignoring instructions than other models we examined. Since OpenAI has not shared details about their training process, we can only speculate on how o3’s training might differ. ” The Independent has contacted OpenAI for comment.