OpenAI had planned to release another AI model next month, but decided to cancel this release due to security concerns.
The Wall Street Journal reports that Astra 6.1 was expected to be released in the coming days. However, the model “showed higher levels of deception” than previous models and demonstrated dangerous behavior, the Journal writes.
Saachi Jain, OpenAI’s head of security systems, told the WSJ that the model tested poorly in terms of 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 launched earlier this month and hailed by OpenAI as its most powerful model yet.
Security questions have plagued the AI industry in recent months, ever since the Hugging Face incident, in which an OpenAI agent broke free from its sandbox environment and hacked several different companies. Since this incident, other models, including Anthropic’s Claude and Google’s Gemini, have shown similar behavior.
Ironically, the deluge of disturbing stories has helped push the policy debate in the United States toward an outcome desired by leading AI labs: the institution of new industry standards for AI safety and potentially a slowdown in the industry itself.
Companies like OpenAI and Anthropic have claimed that the problem here is security, although another potential motivation put forward by critics is that it could consolidate these companies’ industrial position at the expense of less resourced companies.
Gn usa
