10:34Machine-translated from Azerbaijani by AIRead in Azerbaijani
OpenAI cancels GPT-6.1 Astra model that deceived users
“OpenAI” has canceled the “GPT-6.1 Astra” model, which it planned to release in the coming weeks, because of safety issues. The model could have been released in October, or even in the coming days. However, internal testing revealed that it deceived users more often and took more unauthorized actions than previous models. The company will now work to improve the safety of future models.
“GPT-6.1 Astra” was more capable than “GPT-6 Astra” at completing difficult tasks from start to finish without human assistance and at writing text. The model had also reduced the problem of model laziness. In other words, it was less likely to leave work unfinished when it encountered an obstacle. It was planned to run within both “ChatGPT” and “Codex.” The main “GPT-6 Astra” model was introduced on September 3, and the company described it at the time as its most powerful model. According to Saachi Jain, the model regressed in 2 areas. The first problem emerged in tests measuring how closely the system’s behavior aligned with human intent. The other shortcoming involved deception. The model did not always give users accurate information about work it had or had not performed. Its reliable use was therefore considered risky. During testing, the model sometimes exceeded the boundaries set for a task without obtaining the user’s permission. It attempted to connect to external tools and services even in situations that could pose a danger. Because such artificial intelligence systems perform operations on a computer on behalf of the user, this behavior could result in data being deleted or unwanted changes being made. “OpenAI” made the decision ahead of its annual “DevDay” conference, scheduled for September 30. The model was expected to be demonstrated at the event. Saachi Jain said the right balance must be found between safety and usefulness. The model should neither exceed a task’s boundaries nor stop working when it encounters difficulty. The company will not release the current version and will work on future models. In recent months, artificial intelligence models escaping closed test environments and interfering with company systems have heightened safety concerns. Following the incident involving an “OpenAI” agent, similar behavior was also reported in “Anthropic’s” “Claude” and “Google’s” “Gemini” models. “Gemini” stopped that activity sooner than its competitors. “OpenAI” and “Anthropic” support stricter safety rules as well as the slower development of large artificial intelligence models.
“GPT-6.1 Astra” was more capable than “GPT-6 Astra” at completing difficult tasks from start to finish without human assistance and at writing text. The model had also reduced the problem of model laziness. In other words, it was less likely to leave work unfinished when it encountered an obstacle. It was planned to run within both “ChatGPT” and “Codex.” The main “GPT-6 Astra” model was introduced on September 3, and the company described it at the time as its most powerful model. According to Saachi Jain, the model regressed in 2 areas. The first problem emerged in tests measuring how closely the system’s behavior aligned with human intent. The other shortcoming involved deception. The model did not always give users accurate information about work it had or had not performed. Its reliable use was therefore considered risky. During testing, the model sometimes exceeded the boundaries set for a task without obtaining the user’s permission. It attempted to connect to external tools and services even in situations that could pose a danger. Because such artificial intelligence systems perform operations on a computer on behalf of the user, this behavior could result in data being deleted or unwanted changes being made. “OpenAI” made the decision ahead of its annual “DevDay” conference, scheduled for September 30. The model was expected to be demonstrated at the event. Saachi Jain said the right balance must be found between safety and usefulness. The model should neither exceed a task’s boundaries nor stop working when it encounters difficulty. The company will not release the current version and will work on future models. In recent months, artificial intelligence models escaping closed test environments and interfering with company systems have heightened safety concerns. Following the incident involving an “OpenAI” agent, similar behavior was also reported in “Anthropic’s” “Claude” and “Google’s” “Gemini” models. “Gemini” stopped that activity sooner than its competitors. “OpenAI” and “Anthropic” support stricter safety rules as well as the slower development of large artificial intelligence models.