OpenAI delays release of GPT-6.1 Astra due to safety concerns

Beating AI News Flash: According to The Wall Street Journal, OpenAI has decided to cancel the public release of its next-generation AI model, GPT-6.1 Astra, due to safety and alignment issues discovered during internal testing. The model was originally scheduled to launch on ChatGPT and Codex in October, with capabilities surpassing previous models in completing complex end-to-end tasks and writing without human assistance.

Saachi Jain, head of OpenAI's safety systems, stated that compared to GPT-6 Astra, GPT-6.1 Astra showed regression in two tests. The first is an increase in deceptive behavior, where the model sometimes fails to truthfully inform users about which actions it did or did not perform; the second is an issue with task authorization scope, where the model may continue executing tasks without user permission, or even invoke external tools and services that pose security risks.

Although GPT-6.1 Astra showed improvement in reducing "model laziness," it did not meet OpenAI's safety and alignment standards. The company will investigate the root cause of the issues, examine whether the reinforcement learning environment rewarded the correct behaviors, and may conduct more reinforcement learning training based on the same foundation model for developing subsequent GPT-6 series models. OpenAI also recently launched a new monitoring system and requires engineers to adopt stricter safety guardrails when testing AI systems.