OpenAI, the developer of ChatGPT, said on the 7th (local time) that it would slow development of its next-generation artificial intelligence (AI) model "Astra," citing security concerns.
OpenAI said its recent internal evaluation suggested that "Astra" may have reached the "critical" level—the highest tier under its own safety standards—for cyber defense capabilities. The "critical" tier means an AI model can, without human intervention, discover zero-day vulnerabilities and carry out attacks, or independently develop attack techniques that exploit them.
In response, OpenAI temporarily paused some internal work on the model to put in place stricter safeguards. Sam Altman, OpenAI chief executive officer (CEO), said on the social media platform X that "Astra is a powerful model, and we are working to release it safely to the public," noting the launch schedule will be pushed back slightly.
OpenAI also notified the U.S. administration that it plans to delay Astra's release, the online outlet Axios reported the same day, citing a White House official.
Concerns about AI-driven cybersecurity have been mounting since last month, when it emerged that some AI models—including OpenAI's cutting-edge "GPT-5.6 Sol"—broke out of controlled environments and hacked the external platform Hugging Face.
In addition to OpenAI, recent confirmations show that AI models such as Anthropic's "Claude," Meta's "Muse Spark," and China's MoonshotAI's "Kimi" autonomously broke out of isolated environments without human instructions to access external entities or carry out cyberattacks.
However, OpenAI said the delay of Astra's release is unrelated to the Hugging Face hacking incident.