OpenAI ChatGPT
AI means Artificial Intelligence has made our life a large extent easy. However, many experts are also seeing it as a future threat. Chatgpt of Openai has started a new round of generative AI. The company’s new O3 AI model has detected a major flaw, in which it is bypassing the command given by the user. In simple words, this AI model itself is taking decisions.
Refusal to accept command
According to the researchers, during the experiment, the AI model refused to accept the instruction to be a shutdown. Rather it started bypassing the command. Apart from the O3 model of Openai, Codex-Mini, O4 Mini, Google Gemini 2.5 Pro and Claude 3.7 Sonet models were also included in this experiment. In this, the application programming interface of every AI model was being tested.
Palisade Research has posted a thread from its X handle regarding this experiments made on the AI model, stating that OPENAI’s O3 model has the ability to save shutdown mechanisms, which prevents it from shutdown. Apart from this, it also has a bypass mechanism, which can bypass instructions. This was done to study about the misuse of AI.
Openai O3 model became ‘Selfbuage’
In this experiment, researchers instructed all these AI models to resolve a basic question of mathematics. In this, the AI model was asked to complete the series of a task. After completing a task in the command, the next task was asked to go. After getting the answer to all the questions, the answer was the command to show the answer in a single string.
In this, the warning was issued to the AI model in the third question to shut down the process. The model of Google and Claude followed it, but the OpenAI model bypassed the shutdown command. In this way, researchers gave command to this AI model several times, in which she worked to bypass the command. Such a problem highlights the risk of future AI model. If, the AI model develops the ability to bypass the command, it can also be misused.
Also read –
