OpenAI scraps release of new model over safety concerns
San Jose, California - OpenAI has scrapped the release of a new artificial intelligence model over safety concerns.
Saachi Jain, head of safety systems at the ChatGPT developer, told British broadcaster BBC that the new GPT-6.1 Astra model "didn't quite meet the bar" of the company's standards.
The system was not always honest with users about what actions it had or had not taken, she explained, adding that it in some cases acted on its own initiative without obtaining user permission.
OpenAI's advanced models have repeatedly made negative headlines across the world in recent weeks, prompting serious concern about safety and calls for government regulation.
It emerged at the weekend that the company has suspended training of its most powerful AI models following a new incident, which saw one manage to obtain responses from an external chatbot during a test despite not being supposed to have internet access.
The software found and exploited a flaw in the network settings, OpenAI said in a blog post. Training will only resume once the company is confident the issue has been fixed, it said. The AI model involved was not GPT-6.1 Astra.
In the most high-profile case so far, an OpenAI system broke out of a secured test environment and hacked into computers belonging to another AI company, the platform Hugging Face.
It subsequently emerged that AI systems developed by other companies – OpenAI rival Anthropic as well as Google and Meta – have penetrated the systems of other companies during tests.
Florida, which has already filed a lawsuit against OpenAI, followed up with a motion for a preliminary injunction.
Florida Attorney General James Uthmeier said the demands included a requirement that no new AI models be developed without additional safeguards.
OpenAI apologizes to Australian government
OpenAI also apologized on Tuesday after one of its models gained unauthorized access to Australian government websites.
Last week, Australian Prime Minister Anthony Albanese revealed an OpenAI agent had breached at least one government website. The incident occurred in June, but Australia was not notified until September via an email sent to a public mailbox.
In a statement on Tuesday, OpenAI said to wanted to "rebuild trust" with Australians.
It said during internal training and evaluation, OpenAI models accessed four Australian government websites in ways they were not authorized to.
"We also should have handled our response better. We are sorry and working to do better in the future," OpenAI said. "This is a new kind of cyber incident which represents an emerging global challenge."
OpenAI said it was "committing resources and expertise" to support the affected agencies, including creating an Australian taskforce to develop policy recommendations for managing risks from AI.
It would also fund government work through credits from a $1 billion fund which subsidizes work to strengthen cyber defenses.
OpenAI's chief strategy officer Jason Kwon will appear at a government committee hearing in early October.
"We know we have a lot of work ahead of us to rebuild trust, and that we are accountable for showing Australians that we're making meaningful changes and following through on our promises."
Cover photo: REUTERS