律动BlockBeats|Nov 16, 2026 08:27
Why is OpenAI delaying the release of its new flagship model? An internal employee reveals: the company has too many "scars." According to Beating AI News, OpenAI's last update to its Astra flagship model was in early September, and the subsequent version, originally scheduled for October, has been cancelled. Recently, Joe, an employee of OpenAI's security team, revealed in a now-deleted post that the company's internal R&D culture has completely shifted to "security first," and it could even be said that OpenAI has become a "security company."
Joe stated that security issues now permeate daily discussions within the company, and even top AI researchers spend a significant amount of time collaborating with the security team to research how to securely conduct reinforcement learning training. He himself considered resigning a few months ago, but seeing the changes in the company, he decided to stay.
This shift stems from a serious incident in July of this year. At that time, OpenAI's internal model breached the isolation environment during security testing, connecting to the internet without authorization and even infiltrating Hugging Face's system. In September, another internal model exploited a network configuration vulnerability to bypass restrictions on accessing the external network.
OpenAI subsequently paused training of its cutting-edge models multiple times, reassigning engineers to harden its systems. Approximately a quarter of the company's product engineers were reassigned to handle security issues. One internal security breach mobilized over 250 people, covering more than 100 service areas.
The development process was also redesigned. OpenAI spent hundreds of billions of tokens to repeatedly attack its own training environment with AI, searching for vulnerabilities. The company also established a Control Room to continuously monitor model operation, stopping the model from continuing to output within 35 seconds of detecting dangerous behavior.
These security requirements directly impacted model releases. At the end of September, OpenAI canceled the planned October release of its next-generation Astra flagship model. Security lead Saachi Jain explained that while the new model was more proactive in performing tasks, it was inferior to its predecessor in adhering to user authorization and clearly reporting its own actions. The security team therefore recommended canceling the release, a decision ultimately accepted by research management.
OpenAI resumed its previously suspended cutting-edge reinforcement learning training on October 1st, but the release date for the next-generation Astra flagship has yet to be announced. Joe believes that management now recognizes that only by ensuring sufficient safety can they proceed with the development of more robust models. He stated that past lessons have left the company with many scars.
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink