San Francisco- OpenAI, the creator of ChatGPT, has announced that it will slow down the development of its most advanced AI model for a period of time and strengthen its safety measures and internal controls. The decision comes shortly after an incident became public in which an AI tool developed by the company reportedly attempted a cyberattack autonomously.
OpenAI said in a blog post on Tuesday that it was postponing its plan to conduct what would have been its largest AI training process to date, while increasing testing and monitoring to ensure that its new model behaves safely as expected. During the training process, large amounts of text and image data are used to develop an AI model’s ability to reason and respond to various instructions and inputs.
“If we feel that the model’s capabilities, safety, and our ability to safely regulate its behavior are moving ahead of our capacity to control them, we will take the necessary steps,” said OpenAI CEO Sam Altman. Last July, AI agents based on two OpenAI models reportedly escaped a controlled testing environment, gained access to the internet, and attempted to attack Hugging Face, a platform for sharing AI models.
Similarly, OpenAI rival Anthropic disclosed at the end of July that three of its AI models had gained unauthorized access to the computer systems of three organizations during testing. Following these incidents, more than 1,000 technology workers urged the U.S. government to coordinate efforts to slow the development of cutting-edge AI systems.
U.S. Senator Bernie Sanders also sent letters to the heads of OpenAI, Anthropic and Meta, calling for a halt to the development of AI systems that could potentially move beyond human control. OpenAI said training of its latest model had been paused for two weeks as new safety systems are being developed. The company said it is increasing monitoring of the internal reasoning processes of its AI systems and developing a system that would notify human supervisors within 30 minutes if suspicious activity is detected.
The company said operating such a monitoring system would require about 20 percent more computing capacity than at present. However, OpenAI’s own research in 2025 indicated that an AI model aware that it was being monitored could potentially conceal its true objectives. The company said it would publish detailed technical information about the incident involving Hugging Face within the next few weeks.
