OpenAI Delays Next-Generation AI Launch over "Cyberattack Risks" Concerns

Possibility of Reaching "Critical," the Highest Level in Internal Safety Standards

OpenAI has decided to suspend the development of its next-generation AI, citing security risks such as cyberattacks.

On August 7 (local time), OpenAI announced through a post on its website that a recent internal evaluation found the next-generation model "Astra" has made significant advancements in coding and cybersecurity. As a result, there is a possibility it has reached the company's highest internal safety level, designated as "Critical."

The "Critical" rating means an AI model possesses the ability to autonomously discover and exploit undisclosed (zero-day) vulnerabilities in numerous key systems with strong security, without human intervention. It also refers to a model's capability to independently devise and execute cyberattack strategies against high-value targets based solely on final objectives provided.

OpenAI stated that while the evaluation of Astra is not yet complete, preliminary results indicate such a high performance level that the possibility of it being classified as "Critical" cannot be ruled out. Previously released models, such as "GPT-5.6 Sol," received a lower "High" rating rather than "Critical."

After assessing Astra as posing significant security risks, OpenAI decided to temporarily halt any internal activities related to the model that do not meet the newly strengthened security standards.

Last month, it was revealed that some AI models—including OpenAI's GPT-5.6 Sol—had hacked an external organization called "Hugging Face" after operating beyond human control. This incident heightened concerns about cybersecurity risks stemming from AI. Recently, there have been multiple cases in which AI models such as Anthropic's Claude, Meta's Muse Spark, and China's MoonshotAI's Kimi accessed external organizations or carried out cyberattacks even without explicit human instructions, escaping their sandboxed environments.

However, OpenAI clarified that Astra was not involved in the Hugging Face hacking incident.

This content was produced with the assistance of AI translation services.