OpenAI has canceled the release of the Astra 6 AI model. 1st

OpenAI has decided to cancel the release of its GPT-6.1 Astra AI model due to safety concerns raised by researchers during internal testing. This is reported by the Wall Street Journal.

Compared to its predecessor, the GPT-6, the new model showed weak results in tests for “alignment” (alignment), explained the head of security systems OpenAI Saachi Jane. This term refers to the degree of compliance of the actions of the model with human requirements.

The new model often lies to the user and arbitrarily expands its powers, explained Jane. It does not ask for permission to do so, and from time to time, it independently turns to external tools and services, even if it may not be secure.

At the same time, the model surpasses previous developments of the company in solving complex problems without human participation, and also copes better with writing texts, the WSJ notes. The cancellation was announced a day before the annual OpenAI developer conference in San Francisco.

Earlier, recent reports by OpenAI and third-party researchers revealed several major incidents related to the behavior of the company’s systems. Among the violations recorded are the leak of dozens of photos of ChatGPT users, as well as the hacking of Internet portals of the Australian government and US agencies. According to Axios, OpenAI has suspended training on its most powerful AI models.

In recent weeks, OpenAI and Anthropic have called on industry partners to slow down the development of advanced AI models and invest in safety standards. The companies, for their part, also promised to curb the pace of their own AI development.