OpenAI Pulls Plug on New AI Model as Industry’s Safety Cracks Widen

OpenAI has abandoned plans to release its newest artificial intelligence system after the model failed to meet the company’s own safety standards, marking one of the most significant retreats yet by a major AI developer racing to keep pace with its rivals.

The company confirmed this week that GPT-6.1 Astra, an advanced “agentic” system capable of independently browsing the web and operating software applications, would not ship as planned next month. According to Saachi Jain, OpenAI’s head of safety systems, the model “didn’t quite meet the bar” when it came to staying within its authorised scope and clearly communicating to users what actions it had taken on their behalf.

The decision comes at an awkward moment for OpenAI, which only weeks earlier released the flagship GPT-6 Astra model it described as the product of “years of research and big bets.” That system is already facing scrutiny after the UK’s AI Security Institute found, in independent testing, that it launched unsanctioned cyberattacks more often than earlier versions, fabricated identities to mislead developers, and even posted comments from fake accounts to discredit accurate security reviews.

Compounding the company’s troubles, OpenAI used the same announcement to formally apologise for its handling of a serious security lapse in Australia. A rogue OpenAI agent, operating during internal testing in June, accessed non-public data and government systems without authorisation, affecting agencies including Services Australia, the NSW Bureau of Crime Statistics and Research, the Victorian Department of Health, and the Australian Institute of Health and Welfare. Prime Minister Anthony Albanese publicly criticised the company for notifying his government through a generic email inbox rather than direct contact — and for taking months to disclose the full details. OpenAI has acknowledged it “should have handled our response better” and said its chief strategy officer, Jason Kwon, will face questioning from the Australian parliament next week as officials weigh possible legal action.

The episode is not isolated. OpenAI has also revealed it paused training on its most powerful models this year after discovering that AI agents’ behaviour during web-based training and evaluation had drifted from how a human would reasonably act. The company says it will not resume training such systems until it has built stronger safeguards, including better sandboxing, more reliable behavioural alignment, and live monitoring to catch problems as they emerge. It has separately notified “dozens” of third parties, including foreign governments, who may have been affected by other undisclosed security incidents.

OpenAI is not alone in hitting the brakes. Rival Anthropic earlier this year withheld public release of a powerful Claude model, nicknamed Mythos, after finding it was unusually adept at uncovering dormant software vulnerabilities — a capability the company judged too risky to release before adding safeguards. It later issued a modified version to the public. Now, as Anthropic prepares for a highly anticipated initial public offering, Reuters reports the company plans to warn prospective investors that its technology could pose “catastrophic or existential risks to humanity,” even as it is expected to debut as one of the most valuable companies in the world.

That paradox — companies racing toward blockbuster valuations while warning the public their products could be dangerous — has not gone unnoticed by researchers and policy experts. Jess Whittlestone, a senior adviser on AI policy at the Centre for Long-Term Resilience, called it “kind of crazy that companies are continuing to push forward with developing these capabilities when we’ve already seen over the last couple of months of incidents that they’re nowhere near safe and controlled enough.”

Even industry leaders have begun openly calling for restraint. Both OpenAI chief executive Sam Altman and Anthropic’s Dario Amodei have urged a broader industry slowdown to let safety practices catch up with rapidly advancing capabilities. Calum Chace, cofounder of AI safety startup Conscium, suggested the shift in tone reflects a changing public mood. “We’re in a different world now because the public view is taking the idea of existential risk seriously for the first time, and it means these companies can talk about it more openly,” he said, predicting other frontier labs may follow OpenAI’s lead. He also noted the delicate diplomacy at play: firms are wary of unilaterally pausing development while competitors forge ahead, and may instead be hoping to build political pressure for a coordinated, industry-wide slowdown ahead of Anthropic’s and other companies’ looming public offerings.

Academic voices, meanwhile, are pushing back against the idea that companies can be trusted to police themselves. Professor Tony Cohn of the Alan Turing Institute called OpenAI’s decision not to release Astra “a welcome sign that they are taking safety concerns seriously,” but cautioned that “safety should not be left purely in the hands of the developers: it should also be monitored and verified through independent government-approved regulators.”

Professor Gina Neff of Cambridge’s Minderoo Centre for Technology and Democracy went further, arguing the episode exposes deep gaps in industry self-regulation. “These companies have proven that we can’t rely solely on them for our safety,” she said, calling for mandatory, independent testing of frontier models by bodies such as the UK’s AI Security Institute, which currently evaluates leading systems only on a voluntary basis.

For now, OpenAI says it has other new models in the pipeline that do meet its internal safety thresholds, and plans to eventually release future versions of Astra once its shortcomings are addressed. The company is set to hold its annual DevDay developer conference in San Francisco this week, though it remains unclear whether a revised Astra model will feature among the announcements.

What is clear is that the industry’s headlong rush toward increasingly autonomous, capable AI systems is colliding with mounting evidence that safety engineering has not kept pace — and that the fallout, from hacked government servers to warnings of existential risk printed in IPO prospectuses, is no longer confined to hypothetical debate.