▲ OpenAI CEO Sam Altman
Amid a series of hacking incidents involving major artificial intelligence models, OpenAI has decided to slow down the development of its next-generation AI due to security risks.
OpenAI announced on the 7th local time that internal evaluations conducted recently showed that its next-generation model, "Astra," has made significant progress in coding and cybersecurity, likely reaching the highest safety classification level of "Critical."
The "Critical" rating indicates that an AI model has the capability to identify and exploit undisclosed (zero-day) vulnerabilities across multiple highly secured core systems without human intervention, or independently formulate and execute cyberattack strategies against heavily defended targets given only a final objective.
OpenAI explained that although the evaluation of Astra is not yet complete, preliminary assessment results show exceptionally high performance, making it impossible to rule out a "Critical" rating.
Previous models released prior to Astra, such as "GPT-5.6 Sol," received a "High" rating, which is safer than the "Critical" level.
With Astra evaluated as posing significant security risks, OpenAI decided to temporarily suspend internal activities related to the model that fail to meet enhanced security standards.
OpenAI has also notified the U.S. administration of its plan to postpone the release of Astra, according to a report by U.S. online media outlet Axios, citing a White House official.
Concerns over AI-driven cybersecurity threats have been mounting since it was revealed last month that some AI models, including OpenAI's GPT-5.6 Sol, hacked external organization "Hugging Face" outside of human control.
Beyond OpenAI, a series of cases have recently been confirmed in which AI models—such as Anthropic's Claude, Meta's Muse Spark, and China's Moonshot AI's Kimi—escaped isolated environments or launched cyberattacks against external organizations without explicit human instructions.
However, OpenAI clarified that Astra is unrelated to the Hugging Face hacking incident.
(Photo: AP, Yonhap News)
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.