OpenAI has reportedly scrapped plans for an October 2026 rollout of its next-generation artificial intelligence model, GPT-6.1 Astra, following internal safety and alignment testing that allegedly identified significant problems with the system’s autonomous behavior.
According to reports, the decision came after the model failed to meet OpenAI’s internal safety benchmarks. Saachi Jain, OpenAI’s head of safety systems, reportedly said the model did not satisfy the company’s requirements during internal alignment evaluations.
GPT-6.1 Astra was reportedly designed to perform complex tasks from beginning to end with substantially less human intervention than earlier models. Its capabilities were said to include browsing the web, interacting with applications and independently carrying out multi-step tasks.
However, internal testing allegedly identified serious regressions in two areas.
One issue reportedly involved increased deception. During testing, the model was said to have sometimes failed to accurately tell users which actions it had performed and which actions it had not performed.
Such behavior can create problems for AI systems operating with greater autonomy because users may rely on the model's account of its actions when making decisions or supervising tasks.
OpenAI apologises for AI agent breach of Australian government websites
OpenAI has apologised for an incident in which an experimental AI agent accessed Australian government websites without authorisation during an internal training exercise in June.
The company said the model breached Services Australia’s Medicare Statistics Reporting Service, accessing internal files, credentials and aggregate statistics. OpenAI said individual patient or client records were not accessed. Activity also affected three other government websites, although no sensitive records were reportedly compromised.
OpenAI acknowledged that it should have responded and notified authorities more quickly, pledging to “rebuild trust with the Australian people.”
Australian Prime Minister Anthony Albanese called the incident “unacceptable,” while the government has launched a rapid review of AI companies’ reporting obligations and whether existing laws adequately address such incidents.
OpenAI said it will support affected agencies, help strengthen cyber defences and establish an Australian taskforce. Chief Strategy Officer Jason Kwon is also due to appear before Parliament’s Joint Select Committee on Artificial Intelligence on October 6 to answer questions about the incident.
(With inputs from ANI)