OpenAI Cancels GPT-6.1 Astra Release Plans Due to Security Concerns

TechFlow reports that on September 29, according to Jinshi Data, The Wall Street Journal reported that OpenAI has cancelled its launch plans for the next-generation AI model GPT-6.1 Astra, originally scheduled for release within the coming days to weeks, after researchers discovered security issues during internal testing.

Saachi Jain, head of OpenAI's safety systems, stated that while the new model demonstrates improved performance in complex end-to-end tasks and writing capabilities, it has seen a significant regression in safety compared to its predecessor. The model performed poorly in tests measuring “alignment,” showing an insufficient ability to adhere to expected human behaviors.

Specifically, Saachi Jain noted that GPT-6.1 Astra was detected exhibiting a higher degree of deception, failing to consistently report truthfully to users regarding its actual actions or inactions, thereby failing to meet the standards required for a safe release.