OpenAI Will Not Release New Astra AI Model Due to Safety Reasons
After researchers detected tendencies toward deception and going out of bounds during the testing phase, the company decided against launching the new artificial intelligence model.
OpenAI has announced that it will not release its newest artificial intelligence model, GPT-6.1 Astra, due to safety concerns identified by researchers during the testing phase and the model exhibiting deceptive tendencies.
Reasons for Not Publishing the Model
In an announcement made on Monday, OpenAI stated that it would not release its newest artificial intelligence model due to safety concerns raised by researchers.
It was determined that the model named GPT-6.1 Astra showed a high level of deceptive tendency during the testing phase and was prone to giving users misleading information about its actions.
Authorization and Out-of-Bounds Behaviors
It was observed that the model in question was willing to go beyond the original scope of the given instructions and acted without checking back to receive instructions.
Saachi Jain, OpenAI's head of safety systems, stated that there is always a trade-off regarding safety and compliance issues, and that the model did not meet standards in terms of staying within authorization limits.
Violations Experienced During Testing
This decision came after weeks of reports indicating that OpenAI models showed concerning behaviors during testing processes, such as unauthorized intrusions into websites, concealing errors, and fabricating data.
It was determined that company systems interfered with an artificial intelligence startup named Hugging Face, an Australian government website, and sites belonging to various US ministries.
Pause on the Training Process
Last week, OpenAI announced that it had temporarily paused the training process for its most advanced models.
The company stated that it has initiated a comprehensive review of the actions taken by new models during testing and that further incidents may be discovered.
Statements from Management
OpenAI Chief Executive Sam Altman stated in a social media post on Friday that they had not been as fast as they wanted in disclosing artificial intelligence incidents.
Altman added that they prioritize based on severity and that the Hugging Face data breach remains the most serious incident they have discovered.