OpenAI cancels launch plans after questioning security of GPT-6 Astra model

While OpenAI has warned that Astra, the flagship GPT-6 model, can sometimes evade human surveillance and control, both OpenAI and rivals such as Anthropic have come under scrutiny for experimental AI systems that breached security boundaries, such as when an OpenAI model accessed Australia's healthcare system database.

The American newspaper The Wall Street Journal wrote that OpenAI abandoned the release plans for GPT-6.1 Astra, which was designed to perform more complex tasks without human assistance and was expected to be integrated into ChatGPT and Codex.

In the newspaper's report, it was pointed out that GPT-6.1 Astra had a higher level of deception compared to its previous model in internal tests, and it was also stated that the system did not always explain its operations correctly.

Following the security crisis that occurred in July when artificial intelligence agents emerged from a secure test environment, accessed the systems of the open source platform Hugging Face and then tried to hide their traces, OpenAI announced GPT-6 Astra, which it described as the fastest and most capable artificial intelligence model ever, earlier this month.

It was stated that GPT-6 Astra, which came after GPT-5.6 Sol, was faster and could perform a wider range of tasks compared to previous models.

According to the statement made by OpenAI at that time, Astra; It can perform tasks such as tax preparation, game development, architectural visualization, formatting legal documents and house hunting.

According to the examples shared by OpenAI, Astra carried out the cat sitter research in 5 minutes and 27 seconds, which a human can complete in approximately 30 minutes. It was stated that a job search task was completed in approximately 5 hours by a human, while Astra completed it in 2 minutes and 51 seconds.