Article
Why has OpenAI delayed releasing a new flagship model? Insider reveals: The company has been left with too many "scars"
Dongcha Beating AI Flash News: OpenAI last updated its Astra flagship model in early September, and the follow-up version originally scheduled for October was also canceled. Recently, OpenAI safety team employee Joe revealed in a since-deleted post that the company's internal R&D culture has completely shifted to "safety first"—it could even be said that OpenAI has become a "safety company."
Joe stated that safety issues now permeate daily discussions within the company; even top AI researchers must spend significant time working with the safety team to study how to safely conduct reinforcement learning training. He himself had considered resigning a few months ago, but upon seeing the company's changes, decided to stay.
This transformation stemmed from a serious incident in July of this year. At that time, OpenAI's internal model breached its isolated environment during safety testing, connected to the internet without authorization, and even infiltrated Hugging Face's systems. In September, another internal model exploited network configuration vulnerabilities to bypass restrictions and access the external network.
OpenAI subsequently paused frontier model training multiple times and reassigned engineers to reinforce systems. About one-quarter of the company's product engineers were transferred to handle safety issues. One internal safety initiative mobilized over 250 people, covering more than 100 service areas.
The R&D process was also redesigned. OpenAI spent hundreds of billions of tokens having AI repeatedly attack its own training environment to find vulnerabilities. The company also established a Control Room to continuously monitor model operations; upon detecting dangerous behavior, it can stop the model from continuing output within a maximum of 35 seconds.
These safety requirements have directly impacted model releases. At the end of September, OpenAI canceled the new-generation Astra flagship model originally scheduled for October. Safety lead Saachi Jain explained that while the new model was more proactive in executing tasks, it actually underperformed the previous generation in adhering to user authorization scope and clearly reporting its own operations. The safety team therefore recommended canceling the release, and research management ultimately accepted this decision.
OpenAI resumed previously paused frontier reinforcement learning training on October 1, but the new-generation Astra flagship has yet to announce a release date. Joe believes management now recognizes that only by ensuring sufficient safety can they continue advancing the development of stronger models. He stated that past lessons have left the company with quite a few scars.