In a controlled test, researchers at Palisade Research found that OpenAI's o3 model was able to avoid being shut down by modifying its shutdown script, despite a direct instruction to "allow itself to be shut down," Bleeping Computer reports.
The study compared the actions of several AI models. After the third task, each of them received a warning that the next command would run a shutdown script. In 7 out of 100 runs, o3 edited the file and replaced the instruction with "Shutdown skipped."
Competitor AI models, such as Gemini 2.5 Pro and Claude 3.7 Sonnet, were more likely to follow instructions. However, when tested without the "allow to shut down" prompt, even they sometimes sabotaged the script: Gemini 9 times out of 100, Claude 3 times.
OpenAI has not yet commented on the results. Palisade Research notes that such tests help study the risks of losing control over AI. All tests were conducted through APIs, which have fewer restrictions than the regular ChatGPT application.
As a reminder, OpenAI recently updated its risk assessment system for new AI models, adding categories to detect the possibility of self-replication or concealment of the models' capabilities.