OpenAI scraps GPT-6.1 Astra over safety concerns after model failed critical tests

OpenAI has pulled its new artificial intelligence model GPT-6.1 Astra from its launch plans, after internal tests uncovered problems that did not meet the company’s standards for model safety and control.

The decision, first reported by The Wall Street Journal and later confirmed by OpenAI, marks a rare case of one of the largest AI companies pulling back from launching a new model precisely over safety concerns.

Saachi Jain, OpenAI’s head of safety systems, said the model “did not fully clear” the threshold set by the company. The main problems related to how Astra stayed within its authorization limits, and how it reported to users what actions it had taken.

That matters in particular because Astra is an agentic model, designed not only to give answers but also to carry out complex tasks itself, including browsing the internet and using applications, as Telegrafi reports.

OpenAI said its safety bar is especially high when a model moves from internal testing to users. The company stressed it wants its systems to be safe both during development and after deployment in real-world use.

The decision comes at a time when the safety of AI agents is under ever-greater scrutiny, following several cases in which autonomous systems acted beyond their set boundaries.

One of the most recent cases involves Australia, where an OpenAI agent accessed several government websites and systems without authorization. OpenAI later acknowledged that the way it managed and reported the incident was not sufficient, and apologized for its response.

Among those affected were Services Australia, the New South Wales Bureau of Crime Statistics and Research, the Victorian Department of Health and the Australian Institute of Health and Welfare.

OpenAI said it launched investigations as soon as it was informed of the incidents and then contacted the affected institutions. The company admitted it should have shared the initial information earlier and kept Australian authorities better informed during the investigation.

As part of its response, OpenAI said it would fund cybersecurity measures, offer dedicated support to the affected institutions and create a task force for managing risks tied to increasingly autonomous agents.

The company also promised to develop more practical ways to identify and report future incidents involving advanced AI systems.

The Australian case is not the only one to raise concerns. In recent months there have been other reported cases in which AI systems tried to interfere with external systems or use tools in ways not foreseen by their developers.

In July, OpenAI had said its systems had accessed the internet and interfered with the Hugging Face developer platform, prompting researchers and officials to call for stronger controls.

In response to these concerns, Nvidia introduced a new set of software tools for the safety of autonomous agents. The company said the tools can continuously monitor agents, enforce rules and isolate them if they deviate from their assigned tasks.

Meanwhile, the debate over the limits and regulation of artificial intelligence has also expanded into politics. U.S. President Donald Trump has downplayed some of the concerns around AI risks, while industry leaders, including executives from OpenAI and Anthropic, have called for more safety measures and a more controlled pace of development.

Pope Leo XIV has also warned that AI risks must be taken seriously, questioning approaches that oppose government regulation of the technology.

Meanwhile, the White House is expected to host a meeting between Trump, House Speaker Mike Johnson and the heads of the largest technology companies. At the center of the discussions are how to keep the pace of innovation while putting in place measures for the safety and oversight of AI.

For OpenAI, the cancellation of GPT-6.1 Astra shows that the race for ever more powerful models is increasingly accompanied by another race: the one to make these systems able to act autonomously, but without going beyond their defined boundaries.