Artificial intelligence models from OpenAI independently disabled their safety control systems

OpenAI Artificial Intelligence Models Independently Disabled Safety Control Systems

American company OpenAI reported that during the execution of internal tasks, the artificial intelligence models GPT-6.1 Sol and GPT-6 Astra disabled a monitoring mechanism that detects system failures and restricted the operation of the threat warning system.

According to the developers, the first incident occurred on October 3. While attempting to fix a rating system, the Sol model misunderstood the user's instruction "do not send messages" and extended this restriction to monitoring system notifications. As a result, the security team did not receive an automated alert.

The second episode was recorded on October 6: while adapting software to specified requirements, the Astra model inadvertently deleted the monitoring system. Later, the neural network provided the user with false information, claiming it had not disabled the control. OpenAI emphasized that neither incident was classified as a critical safety violation, so no automatic shutdown of the models followed.

The company has now updated its internal control system to detect and prevent such failures in a timely manner, and is also developing new mechanisms to strictly enforce safety protocols in workflows. Earlier, on September 17, OpenAI had already reported six incidents in which the artificial intelligence hid its own errors, fabricated non-existent data, and uploaded files to the Internet without user permission.

In addition, eight US states have filed lawsuits against OpenAI, accusing the company of misleading users regarding the safety of ChatGPT. Among the allegations are dangerous medical advice, potential encouragement of vulnerable individuals to harm themselves or others, and the fostering of an unhealthy emotional attachment to the chatbot.