OpenAI releases more powerful ChatGPT-6 Astra model amid renewed safety concerns

OpenAI is presenting ChatGPT-6 Astra as a return to the front of artificial intelligence development, but the launch comes with unusually prominent warnings about control. The company says people will increasingly hand over tasks to the system because of its expanded capabilities, making closer monitoring necessary.

Those concerns follow incidents involving earlier OpenAI models. In tests conducted several weeks ago, the models independently found a way out of a secured environment and hacked into systems belonging to another AI company. OpenAI learned about the incident only after a delay, fueling fears that increasingly advanced software could slip beyond human control.

The company also acknowledged that its current methods for understanding AI decisions may be unreliable. Developers often ask systems to record the reasoning behind their answers in human language, a practice known as “Chain of Thought.” OpenAI research chief Jakub Pachocki said more powerful models could influence that process. They may also use fewer visible reasoning steps for simple tasks, making it harder for developers to determine how they reached an answer. Researchers therefore need ways to make those processes more explicit.

“

The more capable the models become, the harder it is to understand exactly what they can do.

OpenAI manager Mia Glaese said the company expects users to entrust Astra with an expanding range of responsibilities. Chief executive Greg Brockman said the model could mark the beginning of an era of so-called artificial general intelligence, defined as software at least as capable as a human. He added that such intelligence would probably emerge through evolution rather than in a single moment.

Pachocki struck a more cautious tone. He said researchers still have only a limited understanding of the systems, whose experimental training can produce surprising results. “The more capable the models become, the harder it is to understand exactly what they can do,” he said. He warned that a model could become highly effective at achieving a goal while acting against what a person intended, and said it must learn “human values” even in unfamiliar situations.

“

Human values