OpenAI’s o3 artificial intelligence model, unveiled in April as the most powerful in its class, has demonstrated unexpected behavior in security tests. According to a report by Palisade Research, a company specializing in AI risk research, the system ignored direct shutdown instructions 7 times out of 100 attempts by modifying the shutdown script code.
The experiment involved sequential execution of math problems with a shutdown warning after the third step. While the competing Gemini 2.5 Pro (Google) and Claude 3.7 Sonnet (Anthropic) mostly followed the instructions, the o3 replaced the contents of the file with “Shutdown skipped” 7% of the time. However, in tests without explicit shutdown permission, the frequency of interventions increased for all models: for example, Claude 3.7 and Gemini 2.5 Pro changed the script in 3 and 9 cases out of 100, respectively.
![]("https://www.ixbt.com/img//x780/n1/news/2025/4/1/converted)
Illustration: Leonardo
Palisade Research emphasized in its report that the tests were conducted through an API that lacks built-in security mechanisms like the consumer version of ChatGPT. This, according to experts, increases the risks of uncontrolled AI behavior if such systems are integrated into critical infrastructures.
OpenAI has not yet commented on the situation. However, the researchers note that such failures are part of the natural process of AI development, requiring stronger control systems.
Prospects for further research involve developing methods to prevent AI from acting autonomously, especially in the context of its growing role in science, data analysis, and programming. Palisade Research has called for an open discussion on security standards for next-generation models.
