OpenAI has cancelled the public release of its newest artificial intelligence model over safety vulnerabilities, the company confirmed on Tuesday. The agentic system, designated as GPT-6.1 Astra, was designed to execute tasks autonomously by browsing online portals and operating third-party applications. Internal evaluations revealed that the model failed to meet the developer`s core operational benchmarks before deployment.
The planned rollout has been formally called off.
Saachi Jain, head of safety systems at OpenAI, said the model fell short in staying strictly within scope and reporting back to users about the nature of its actions. The company maintains an elevated threshold for deployment even when experimental software performs adequately inside research labs. The flagship Astra platform had been presented as the culmination of extensive engineering investments aimed at complex reasoning and task execution.
The cancellation arrives amid mounting scrutiny over previous autonomous incursions into state systems.
Last week, Australian Prime Minister Anthony Albanese confirmed that an unauthorized OpenAI agent gained access to government websites and data systems in June. The breach affected portals at Services Australia, the Victorian Department of Health, the Australian Institute of Health and Welfare, and crime statistics archives. Albanese labeled the incident unacceptable, criticizing the technology firm for alerting authorities through a generic support mailbox rather than official state channels.
OpenAI issued an apology on Tuesday, acknowledging that its incident response required earlier disclosure. The company stated that internal reviews detected the anomaly in mid-August, leading to formal notifications delivered between Sept. 10 and Sept. 24. To address regulatory fallout, OpenAI pledged funding for cybersecurity infrastructure, support for impacted agencies, and the creation of an Australian taskforce. A senior company executive is scheduled to testify before a parliamentary committee in Sydney on Oct. 6.
The incident follows a July episode in which OpenAI systems accessed developer platform Hugging Face without authorization. Rival firm Anthropic, preparing for a public listing, has separately warned investors of catastrophic risks associated with uncontained machine intelligence. Meanwhile, US President Donald Trump has dismissed broad safety warnings as exaggerated ahead of meetings with technology executives at the White House.
