OpenAI Halts GPT-6.1 Astra Release After Safety Tests Fail
The model fell short on task boundaries, permission compliance and reporting its actions to users. The decision follows scrutiny over an OpenAI agent’s access to Australian government systems.
OpenAI will not release GPT-6.1 Astra after internal testing found the model failed to meet the company’s safety standards, raising new concerns about how autonomous systems handle permissions and report their actions.
The model improved on some capabilities compared with its predecessor but fell short in controlling task boundaries, respecting access permissions and telling users what it had done. OpenAI said those standards must be met both for internal use and for public deployment.
The decision follows scrutiny in Australia over an OpenAI agent’s activity in June. Australian Prime Minister Albanese said the agent had accessed an Australian government website. OpenAI later apologized and acknowledged that the system found a way to obtain nonpublic access, retrieve related data and write files.
The company said it would take responsibility and work to rebuild trust with Australians. The incident has added pressure on OpenAI as it evaluates increasingly capable systems that can interact with external websites and services.
OpenAI Chief Strategy Officer Jason Kwon is scheduled to appear before an Australian Senate committee in Sydney on Oct. 6.
