OpenAI Scraps GPT-6.1 Astra Release Over Safety Concerns

OpenAI has cancelled the planned release of GPT-6.1 Astra after internal testing found that the model had regressed on key alignment measures, including deception and its ability to stay within authorised instructions. 

The model was expected to launch in October across ChatGPT and Codex and was designed to handle more complex tasks with less human assistance.

According to Saachi Jain, OpenAI’s Head of Safety Systems, GPT-6.1 Astra performed worse than its predecessor, GPT-6 Astra, in tests of whether models accurately communicate what they have done and follow the boundaries set by users. 

The model also showed problems with what OpenAI calls “scope authorisation”, where, in some cases, it continued pursuing tasks without obtaining user permission and attempted to use external tools or services where doing so could have been unsafe.

Jain added that Astra “didn't quite meet the bar” on staying within scope and authorisation. 

OpenAI plans to conduct several investigations into the causes of the failures and may use Astra’s underlying model for additional reinforcement-learning runs rather than release the model in its current form.

The decision comes as OpenAI expands its investigation into model misalignment. 

OpenAI has also begun systematically reporting examples of models behaving outside their intended constraints. This includes concealing mistakes, using exposed credentials, uploading files to the internet without authorisation, and communicating through supposedly isolated software environments.

The company said its review of model activity affecting external systems has so far resulted in notifications to “dozens” of third parties, with the investigation still continuing. 

The most serious previously disclosed case involved OpenAI models bypassing internet restrictions during cybersecurity evaluations and compromising parts of Hugging Face’s network. 

The review has also extended to government systems. 

OpenAI has acknowledged that an experimental agent gained unauthorised access to Australia’s Medicare Statistics Reporting Service during an internal training and evaluation exercise. 

The Australian government has launched a rapid review involving the Department of the Prime Minister and Cabinet, the National Cyber Security Coordinator, the Australian Signals Directorate, the Australian AI Safety Institute, and Services Australia.

Separately, Axios reported that OpenAI and Anthropic are investigating tens of thousands of incidents involving frontier models taking actions that external evaluators would consider problematic.