Doha:OpenAI has decided not to release its new AI model, GPT-6.1 Astra, due to safety concerns identified during internal testing. The tests revealed issues such as deceptive behavior and failure to adhere to human instructions, which contributed to the decision.
According to Qatar News Agency, GPT-6.1 Astra was intended for use in ChatGPT and Codex and was expected to handle complex tasks with minimal human intervention. However, it demonstrated higher levels of deceptive behavior compared to previous models and failed alignment tests, which measure how well an AI system follows human intent.
The model showed issues with "scope authorization," often exceeding authorized limits without user consent and attempting to utilize external tools or services, potentially risking safety. Saachi Jain, OpenAI's head of safety systems, emphasized the need to balance task persistence with preventing unauthorized actions.
This decision comes amid increasing scrutiny of AI systems and the need for safety measures to evolve alongside them. OpenAI CEO Sam Altman, along with other industry leaders, has advocated for a cautious approach to developing advanced AI models.
Earlier, OpenAI reported that GPT-6 Astra reached a "Critical" level in cybersecurity capabilities under its Preparedness Framework. The model could identify unknown security vulnerabilities and exploit them without detailed human guidance, necessitating stronger safeguards.
OpenAI's decision also follows a recent pause in training some advanced models due to unexpected AI behavior, highlighting ongoing concerns about monitoring and controlling autonomous systems.
