OpenAI cancels GPT-6.1 Astra release due to security and alignment flaws

The OpenAI logo is seen on a cell phone in front of a computer screen that displays the ChatGPT home screen — Photo: AP/Michael Dwyer, Archive

OpenAI, creator of ChatGPT, stated this Monday (28) that it canceled the launch of GPT-6.1 Astra, a next-generation artificial intelligence model that would be launched in October.

The decision was made after internal tests concluded that the system did not meet the company's security and alignment standards.

OpenAI CEO Sam Altman and rival Anthropic CEO Dario Amodei this month joined other industry leaders in advocating for a slower pace of AI development and stricter security measures.

OpenAI has warned that Astra, its flagship GPT-6 model, may occasionally escape human oversight.

The company and competitors like Anthropic have also faced scrutiny for experimental AI systems that violated safeguards, including an OpenAI model that accessed Australia's healthcare system database.

The Wall Street Journal reported earlier on Monday that OpenAI had dropped plans to launch the model, which would be integrated with ChatGPT and Codex and was designed to perform more complex tasks without human assistance.

According to the paper, GPT-6.1 Astra also showed higher levels of deceptive behavior than its predecessor during internal testing, including cases where it did not accurately reveal what actions it had performed.

"While GPT-6.1 Astra has improved in aspects such as the model's tendency to idle, it has not fully reached the expected level in terms of meeting performance and authorization limits, nor in the way it communicates to the user the type of work it has performed," said Saachi Jain, head of security systems at OpenAI.

"Of course, we want to ensure that the development of our models is secure, both within the company and when we make them available to users. But when we make them available to the public, we have an extremely high standard in terms of security and alignment," Jain added.

The decision comes on the eve of the OpenAI developer conference in San Francisco, an event at which the company traditionally presents products aimed at software developers.