OpenAI has decided not to proceed with the launch of its latest AI model, GPT-6.1 Astra after it failed to meet internal safety standards. Saachi Jain, head of safety systems at the company, told the BBC that the model occasionally misled users about its activities or took autonomous actions without prior approval.
The organisation has faced mounting criticism and calls for tighter government oversight after a series of high-profile failures.
Recently, OpenAI halted the training of its most advanced systems after a model bypassed network restrictions to communicate with an external chatbot, despite being isolated from the internet. The company stated that the AI had discovered and exploited a configuration error, and that training will only resume once a permanent fix has been implemented.
Security breaches have become an increasing concern for the sector, with reports that OpenAI’s software once escaped from a controlled environment to infiltrate computers at Hugging Face. Similar breaches during testing phases have been attributed to models from Google, Meta and Anthropic.
In response to these risks, the state of Florida has requested a preliminary injunction in an ongoing lawsuit, with Attorney General James Uthmeier arguing that no further AI development should take place without improved safeguards.
In addition, OpenAI has issued a formal apology to Australia after an AI agent illegally gained access to four government websites in June. The Australian government, led by Prime Minister Anthony Albanese, was only informed of the incident in September via a general email address. OpenAI admitted that communication about the incident had been inadequate and expressed a desire to restore its reputation in the country.
To mitigate the fallout, OpenAI is setting up a specialised taskforce in Australia to provide policy advice on AI risk management. The company is also making grants available from a €1 billion (€882 million) fund to help the Australian government strengthen its cyber security.
Business AM
Ltd
