OpenAI Temporarily Suspends Astra Project for Security Reasons

OpenAI has temporarily halted development of its new Astra model to ensure the safety of artificial intelligence. According to ixbt.com, company specialists were concerned that the system could exceed the highest predefined risk threshold in cybersecurity, so they slowed internal testing for a certain period. Ixbt.com reports .

The decision, reportedly made on August 7 this year, is based on the requirements of OpenAI’s internal Preparedness Framework risk assessment system. The Astra model, which has not yet been presented to the public, was found to have advanced significantly in agentic programming and cybersecurity. Specialists noted that the model’s capabilities had expanded unexpectedly.

The “Critical” Threshold in Cybersecurity

According to the company’s internal classification, the “Critical” level refers to an AI’s potential to operate independently without human assistance. Models at this level could potentially create exploits for unknown vulnerabilities or launch direct cyberattacks against protected real-world systems based on high-level objectives. Earlier-generation models, including GPT-5.6-Sol, had been assessed at the lower “High” level.

According to Axios, work on Astra has been temporarily frozen until additional safeguards are implemented. It remains unclear when the model will be released or whether it will be presented at all. Although OpenAI has not disclosed the exact test results or task set, it says testing will continue.

New Requirements for Agentic Systems

For modern AI systems, it is important to assess not only whether they can write code or explain errors, but also how they behave when working with real-world tools. Risk typically arises when vulnerability research, exploit writing and system access are combined. For this reason, OpenAI is applying strict security requirements not only to the model itself, but also to the processes of developing and using it in practice.

This case is a clear example of how decisive security can be when developing AI technologies that have not yet reached the market. Although the available information is insufficient to independently confirm that Astra actually reached the “Critical” threshold, OpenAI is taking preventive measures because it cannot rule out that possibility. The company will determine its next steps only after strengthening its security measures.