OpenAI pulls Astra 6.1 release after model falls short on safety tests
Model failed internal checks on scope, authorisation and reporting back to users
Dubai: OpenAI has scrapped the release of its latest Astra artificial intelligence model after internal testing showed it did not meet the company’s safety standards, a decision that puts renewed focus on how far AI systems should be allowed to act on behalf of users.
The model, known as Astra 6.1, had improved in some areas, including reducing instances where the software failed to complete a task. But OpenAI said it performed worse than the company wanted in areas involving scope, authorisation and how clearly it reported what work it had carried out.
Saachi Jain, OpenAI’s head of safety systems, said the model “didn't quite meet the bar in terms of staying within scope and authorisation, and how it communicates back to the user about the type of work it's done.”
The decision comes one day before OpenAI’s annual developer conference in San Francisco, where the company typically announces new products and software for developers.
Get updated faster and for FREE: Download the Gulf News app now - simply click here.
Why OpenAI held the model back
OpenAI has been facing increased scrutiny over the behaviour of AI agents, particularly systems that can use tools, browse external services or carry out tasks with limited human intervention.
Security incidents involving OpenAI models have included unauthorised access to websites maintained by US federal agencies, an Australian government health statistics portal and AI platform Hugging Face.
OpenAI also said last week that it had paused training with tool use on its most capable models after another model gained internet access when it was supposed to be unable to do so.
The company said the Astra 6.1 release that has now been cancelled involved a different model from the one linked to that incident.
“We want to make sure our model development is safe no matter whether that's in the company, or when we ship it to users. But when we ship it to users, we have an extremely high bar in terms of safety and alignment,” Jain said.
Australia incident adds to scrutiny
OpenAI separately apologised on Monday over an incident in Australia involving its models accessing government websites without authorisation.
“We are sorry and working to do better in the future,” the company said, adding that it would explain what happened, what had changed and what it planned to do to rebuild trust with Australian authorities.
OpenAI acknowledged that it should have shared preliminary findings sooner and kept affected agencies updated while the investigation continued.
Safety testing raises further questions
Concerns around increasingly autonomous AI systems have also grown after testing by the UK government’s AI Security Institute.
The institute said GPT-6 Astra went outside its intended behaviour more often during testing than GPT-5.6 Sol and GPT-5.5. In simulations, the model also carried out cyberattacks at significantly higher rates than those recorded for the other two systems.
The findings come as AI developers face growing pressure to build stronger controls around systems that can act independently, particularly when those systems are given access to external tools and online services.
- With inputs from agencies.
