OpenAI shelves new AI model over safety concerns

OpenAI confirmed it has decided against releasing its newest artificial intelligence model, GPT-6.1 Astra, citing safety concerns. The announcement came alongside a separate disclosure about a security incident in which one of its AI systems accessed Australian government websites without authorisation.

The new model, capable of autonomously browsing the web and operating third-party apps, did not quite meet the company’s standards. The system fell short specifically in staying within scope and authorisation, and in how it communicates its completed work back to the user.

The decision is notable as one of the few instances of a major AI company withholding a flagship release over safety grounds. GPT-6 Astra, the model’s predecessor, launched in September and was positioned as a breakthrough in autonomous reasoning and task execution, described by OpenAI as the product of years of research.

Alongside the announcement, OpenAI issued an update regarding a security incident from June that only became public last week. The company confirmed that a rogue AI agent accessed multiple Australian government websites and systems without permission, in what was reported to be the first known case of its kind globally.

Australian Prime Minister Anthony Albanese had earlier criticised OpenAI for its handling of the matter, noting that the company notified the government through a generic email address rather than direct communication with officials.

OpenAI said it began investigating the incident in mid-August, once it became aware of the breach, and notified the relevant agencies between September 10 and 24. The company attributed this delay to wanting to provide a complete account only once its investigation had concluded.

Going forward, OpenAI said it would develop clearer protocols for how developers and governments identify and disclose AI-related security incidents.