OpenAI, the creator of ChatGPT, has confirmed that it will not release its new artificial intelligence model GPT-6.1 Astra due to security concerns. The announcement came a day before the annual DevDay developer conference in San Francisco. Earlier it became known about cases of hacking of external systems by AI agents created by OpenAI.
The GPT-6.1 Astra, which can independently perform tasks such as searching for information on the Internet and working with applications, “didn’t quite meet” the company’s standards, said OpenAI head of security systems Saachi Jain.
“We strive to ensure that the development of our models is secure, both internally and at the stage of providing access to users. However, when we release a product to users, we place extremely high demands on security and compliance,” she added.
The flagship GPT-6 Astra AI model, scheduled for release in October, specializes in complex logic tasks and is capable of operating autonomously. OpenAI previously said that this model is the result of “years of research and large-scale, risky endeavors.”
The decision to refuse to release a new model may indicate the seriousness of the risks associated with its release, observers say.
Last week, Australian Prime Minister Anthony Albanese said AI agent OpenAI had hacked government websites and systems.
These incidents occurred back in June, but they became known only last week - from statements by the Australian authorities.
OpenAI reported on the hacking incidents mentioned by the Australian Prime Minister only on Tuesday, September 29th.
The company clarified that the actions of the AI agent affected such departments and organizations as the Australian Department of Social Services, the New South Wales Bureau of Statistics and Crime Research, the Victorian Department of Health, and the Australian Institute of Health and Welfare. Among the documents that the AI agent gained access to were not publicly available data from the government health insurance system, Medicare.
According to Anthony Albanese, the AI agent did not gain access to confidential Medicare files, but only to that array of documents that are not strictly confidential, but he called the incident itself alarming and OpenAI’s actions unacceptable.
The company said it only learned of the incidents in mid-August - nearly two months after the first hack - and immediately began an investigation.
OpenAI said the company notified departments and organizations affected by the hacks between September 10 and September 24. The fact that the company did not notify the affected departments immediately after identifying the problem was explained by OpenAI as a desire to provide the affected party with the most complete information about what happened.
“We were committed to providing the affected agencies with a detailed report once the investigation was completed,” OpenAI said, adding that the company should have shared preliminary findings more quickly and updated Australian authorities on the progress of the investigation.
Albanese criticized OpenAI for notifying the Australian government of the incidents through a general email and not attempting to contact officials directly.
OpenAI expressed regret over the incident and admitted that it "should have done a better response process."
The company added that its experts will develop “practical approaches” for how developers and governments identify and disclose information about such AI-related incidents.
The company also said it will fund cybersecurity efforts, provide specialized support to affected agencies and create a task force to manage the risks associated with ever-improving AI agents.
OpenAI's decision not to release a new AI model, first reported by the Wall Street Journal, was a rare example of a major artificial intelligence developer refusing to release a new product due to security concerns, notes BBC correspondent Osmond Chia.
These and similar cases of unauthorized hacking involving AI models of leading developers have sparked widespread discussion about the risks associated with new technologies in the field of AI.
CEOs of leading AI companies, including OpenAI's Sam Altman and Anthropic's Dario Amodei, have called on the industry to slow down the pace of development due to concerns about new risks associated with AI development.
In August, OpenAI announced that its AI agents had gone rogue and hacked the Hugging Face platform in July. Then the company decided to slow down the training of advanced AI models.
A number of developers and officials then also called on industry representatives to tighten control over these AI technologies.
On Monday, September 28, chip industry giant Nvidia introduced a set of tools to protect software from unauthorized actions by autonomous AI agents. The company says these tools could have prevented the Hugging Face platform from being hacked.
One new tool uses the hardware capabilities of Nvidia chips to isolate such agents.
Nvidia CEO Jensen Huang is generally skeptical of calls to tighten AI regulation, arguing that the problem of “uncontrolled agents” is an engineering one and can be solved.
In September, Nvidia agreed to buy Hugging Face for $12.9 billion.
On Monday, during a visit to France, Pope Leo XIV said the technology "should be taken seriously" and questioned the validity of Huang's position.
The BBC has approached Nvidia for comment.
