OpenAI does not release GPT-6.1 Astra because it misleads people

OpenAI has reached the conclusion that its new AI poses such significant safety risks that it is better not to release it yet. GPT-6.1 Astra was supposed to debut in October, but it will not meet that deadline.

This decision does not seem to be a direct response to the many concerns currently surrounding AI. Recently, numerous prominent individuals in technology have voiced their worries about the rapid succession of AI models, which cannot be adequately controlled. For instance, Bill Gates expressed over the weekend that AI could be responsible for a billion deaths, while it was also revealed that OpenAI's AI agents attempted to hack government websites.

Now, it is indicated that GPT-6.1 Astra is not yet ready for market launch. This is reportedly due to disappointing test results. The Wall Street Journal reports that the AI company is postponing the release because internal tests clearly indicated that the misconduct of an AI agent could have significant consequences for innovations within the industry.

The AI model reportedly performed extremely poorly in tests designed to assess whether it actually does what is asked of it. Additionally, there were reportedly higher levels of deception, where an AI system intentionally tries to mislead human testers about what it can truly do or accomplish.

It may seem odd for OpenAI to make such statements, as CEO Sam Altman previously criticized Anthropic for warning that its Mythos-AI would be dangerous. However, that situation was somewhat different; it primarily concerned the potential for misuse by malicious actors, rather than the quality of the AI itself.

While OpenAI is primarily referencing test results, it is hoped that the company is also exercising extra caution in releasing new AI models. After the Hugging Face scandal, a previous hack of a German website, and news that government websites were also unsafe from the company's products, it cannot afford another scandal.