OpenAI had planned to launch another AI model next month, but decided to reject the launch due to security concerns.
The Wall Street Journal information that the launch of the Astra 6.1 was scheduled for the next few days. However, the model “displayed higher levels of deception” than previous models and exhibited unsafe behavior, the Journal writes.
Saachi Jain, head of security systems at OpenAI, told the WSJ that the model performed poorly on alignment, a measure of how well the program adheres to human intent.
TechCrunch has reached out to OpenAI for more information and will update the article if they respond.
Astra was released earlier this month and hailed by OpenAI as its most powerful model yet.
Questions about security have plagued the AI industry for the past few months, since then the hugging face incidentin which an OpenAI agent broke free from his sandbox and hacked into several different companies. Since that incident, more models, including Claude de Anthropic and Google Gemini – have been revealed to have exhibited similar behavior.
Ironically, the avalanche of troubling stories has helped push the political conversation in America toward an outcome. desired by the best AI laboratories: the institution of new industry standards for AI safety and potentially a decelerate of the industry itself.
Companies such as OpenAI and Anthropic have stated that the concern here is security, although another potential motivation proposed by critics is that it could strengthen the position of the industry of these companies to the detriment of companies with fewer resources.