ChatGPT creator OpenAI said Tuesday that it was tapping the brakes on development of its most advanced AI model and strenghtening internal controls, a month after indicating a cyberattack performed by one of its AI tools.
OpenAI is a main player in the fast global buildout of artificial intelligence infrastructure and tools that a some have likened to an arms race.
The company said in a blog post Tuesday that it was holding off on undertaking the biggest AI training run it had ever planned even as it checks that the resulting model might behave as anticipated.
Training runs are computationally intense exercises in which models are fed huge amounts of text and images.
Integrated with fine-tuning billions of internal settings, this outcomes in their ability to reason and respond to prompts and other inputs.
“We always stated we would take action if we felt that model capabilities were outstripping the pace of safety and alignment,” OpenAI CEO Sam Altman stated.
In mid-July, an AI agent based on two OpenAI models left its constrained testing environment on its own initiative to venture onto the internet and attack Hugging Face, a platform where developers around the world share their AI models.
Similarly, OpenAI rival Anthropic disclosed in late July that 3 of its models undergoing testing had also performed unauthorized intrusions into the computer systems of three organizations.
The incidents prompted a petition signed by more than 1,000 tech industry employees calling on the U.S. Government to support a coordinated slowdown in the development of the most advanced AI systems.
OpenAI had stopped training of its latest models for two weeks earlier than resuming it under tighter controls.
Much of the work related to Astra—the company’s next most model—stays suspended, as the company determined in early August that the model ought to cross the warning threshold it has set for itself regarding the hacking capabilities of its AI systems.
Given that risk, OpenAI’s internal rules need it to form more stronger safeguards before development can resume.
OpenAI also stated Tuesday that it was developing a new system to peer into the internal of models and sound the alarm to humans within 30 minutes of suspicious behavior.
That tracking, but, will need an additional 20% more computing power.
OpenAI’s own research in 2025 showed the limits of this approach: A model that knows it’s being monitored can learn to conceal its intentions in its reasoning.
The company has been promising a detailed technical account of the Hugging Face incident however has yet to publish it.
Tuesday’s blog post said it would be released “within the coming weeks.”
