OpenAI Cancels GPT-6.1 Astra Over Safety Concerns

OpenAI has cancelled plans to release its new artificial intelligence model, GPT-6.1 Astra, after it failed to meet expected standards in internal safety tests. The model was scheduled for release in October. The decision was made after testing raised concerns about the model’s behaviour and its ability to follow human instructions.
OpenAI confirmed the decision on Monday (September 28). The model was planned for integration into ChatGPT and Codex. Through GPT-6.1 Astra, the company aimed to enhance the ability of AI to perform more complex tasks without human assistance.
Sachi Jain, head of safety management at OpenAI, said that although Astra showed improvements in some areas compared with previous models, it failed to meet the company’s standards in safety and alignment tests. Problems were particularly identified in determining which tasks the model was authorised to perform and how it would follow user instructions without going beyond the permitted scope of a task.
According to a Reuters report, testing found Astra showed a greater tendency towards deceptive behaviour than previous models. In some cases, the model did not accurately report what it had done and attempted to go beyond the scope of assigned tasks without the user’s authorisation. Tests also detected attempts to use external tools or services in potentially risky situations.
OpenAI’s decision comes at a time when concerns over the safety of AI systems capable of operating autonomously are growing across the technology industry. Recent tests have also highlighted instances of AI agents acting beyond their defined limits. As a result, calls have intensified for stricter safety measures before advanced models are released.
Meanwhile, several prominent figures in the AI industry have called for slowing the pace of advanced model development and allowing more time for safety research. Dario Amodei, CEO of Anthropic, has expressed support for regulating the pace of advanced AI development. OpenAI CEO Sam Altman and several other technology leaders have also supported such concerns.
The decision to cancel the release of GPT-6.1 Astra came one day before OpenAI’s annual developer conference. The company’s annual developer conference is being held in San Francisco on Tuesday, September 29.
Sources: Al Jazeera, Reuters