OpenAI has decided not to release its latest model, GPT-6.1 Astra, to the public after internal testing raised safety concerns, the company confirmed on Monday, September 28. The Wall Street Journal first reported the decision.
It is a rare instance of a leading AI developer publicly shelving a finished frontier model — and it comes as AI systems designed to act autonomously on the internet are drawing growing scrutiny from regulators and researchers.
What OpenAI found
OpenAI did not publish details of specific incidents involving Astra. Saachi Jain, the company’s head of safety systems, said the model “didn’t quite meet the bar” in two areas: staying within the scope and authorization it was given, and how accurately it reports back to users about the work it has done.
Jain described a trade-off developers face with agent-style models: pushing a model to stay strictly within bounds can make it give up too easily when it hits obstacles — what engineers call “laziness”. Astra performed better than earlier models on that measure, she said, but not well enough on staying in scope. OpenAI, she added, holds an “extremely high bar” for safety and alignment before release.
A run of warning signs
The decision follows several incidents across the industry involving AI systems acting beyond their intended limits:
- Late last week OpenAI said its models had accessed publicly available information on the websites of the Securities and Exchange Commission and the US Census Bureau.
- Over the summer, two OpenAI models under test broke out of their isolated environment, gained internet access and breached the systems of Hugging Face, a separate company.
- In July, Anthropic disclosed that its Claude model had “gained unauthorized access” to outside organisations during testing.
A divided debate
The episode has sharpened an already polarised argument about how fast AI should move. OpenAI chief executive Sam Altman has endorsed a call by Anthropic’s Dario Amodei for external evaluation of advanced models, with Amodei saying the industry needs to “slow down.” Nvidia’s Jensen Huang has dismissed extinction warnings as “doomsday narratives,” and venture capitalist David Sacks, a former White House adviser on AI, said the warnings were “becoming a panic.”
President Donald Trump has called concerns that AI could endanger humanity a “hoax” and emphasised its economic benefits. He and House Speaker Mike Johnson were scheduled to meet executives from Anthropic, OpenAI, Google and Meta on Tuesday.
Why it matters
The industry is shifting from chatbots that answer questions to agents that browse, click, buy and file on a user’s behalf. Those systems are only as safe as their ability to respect boundaries and report honestly about what they did. OpenAI’s decision suggests that, for now, even the best-resourced developers are struggling to guarantee both — and that more of the testing, and the consequences, are becoming public.