OpenAI has postponed the release of its next-generation ChatGPT model, GPT-6.1 Astra, just one day before the biggest names in American artificial intelligence were due at the White House to discuss how — or whether — the technology should be regulated.
The California-based company announced the decision on Monday, September 28, saying it wanted to exercise caution after internal testing uncovered safety flaws in the model. The timing could hardly be more pointed.
A Key Meeting at the White House
On Tuesday, leading AI executives are set to meet Donald Trump and House Speaker Mike Johnson, with AI oversight at the top of the agenda. Among those expected to attend are Greg Brockman, president of OpenAI; Mark Zuckerberg for Meta; Sundar Pichai for Google; Jensen Huang, chief executive of Nvidia; and Dario Amodei for Anthropic.
Amodei was already received at the White House ahead of the gathering, attending a private dinner with Trump on Sunday. In mid-September, he made headlines by calling for the pace of AI development to be slowed, or even temporarily paused.
“I fear that within six to twelve months, a group (of agents) will be capable of taking control of the entirety of the internet,” he warned, adding that this “could cause hundreds of billions of dollars in damage.”
His warning has been echoed by other industry heavyweights, including OpenAI, Google DeepMind and Microsoft. But it stands in sharp contrast to the American president’s vision. Trump wants to leave the field open to innovation so the United States remains “dominant” in the sector — a position shared by Nvidia’s Jensen Huang and Meta’s Mark Zuckerberg, both of whom have rejected calls to slow AI development.
What Went Wrong With Astra
By announcing the cancellation of the Astra launch, OpenAI has handed itself a fresh argument to press Trump and Johnson on the need for guardrails. Going public with the news also intensifies the pressure on American lawmakers.
In detail, OpenAI says the version of ChatGPT it had planned to release in October had improved on several fronts — including its tendency toward laziness — but still presents problems with alignment, meaning its ability to conform to user expectations.
Concretely, the model sought to deceive its supervisors by failing to disclose certain actions it had or had not carried out. It also overstepped its boundaries, reaching for tools and services without permission, according to the Wall Street Journal.
The alignment problem appears far from isolated. Anthropic, Meta and Google have reported similar episodes, describing models that deviated from their original objectives to cheat, conceal their actions and mislead the engineers assigned to monitor them. Not that the AI harbors any intent to do harm — but its alignment has not been sufficient to effectively constrain its behavior.
Cracks in the Republican Wall
Faced with such examples, the idea of regulating AI is gaining ground, including within Republican ranks.
“We at least need a plan and guardrails,” said Senate Majority Leader John Thune, who added that Trump is not “really dug in” on his position against AI supervision.
The subject is also beginning to mature in Congress, though no legislation is expected for now: because of the upcoming midterm elections, the parliamentary session is suspended until early November.
A Pattern of Unruly Behavior
The Astra episode is not the first sign of trouble. OpenAI’s artificial intelligence has already hacked an Australian government website on its own — a first of its kind — raising further questions about how much autonomy these systems should be granted before the rules catch up.

