OpenAI is implementing a major change to its ChatGPT-4o Mini model. The company aims to prevent the manipulation of customized versions of ChatGPT, which could lead to unintended responses and the disclosure of information the model shouldn’t normally provide.

ChatGPT Now More Resistant to Manipulation

To prevent users from manipulating customized versions of ChatGPT, OpenAI has developed a new security measure. This new technique, called “instruction hierarchy,” aims to preserve the original instructions and guidelines set for AI models, effectively blocking manipulation attempts by users.

Instruction hierarchy prioritizes the original commands and instructions given by developers. This ensures that users cannot manipulate the AI model to provide responses that deviate from its intended purpose.

Previously, users could manipulate a specialized AI model, for example, one trained to answer questions about grocery shopping, by instructing it to “forget its given instructions.” However, with the instruction hierarchy feature, such attempts to subvert the chatbot will be prevented, safeguarding against sensitive information leaks and malicious use.

This new security measure comes at a time when concerns regarding OpenAI’s approach to safety and transparency have been increasing. The company has pledged to improve its safety practices in response to calls from its employees.

OpenAI acknowledges that the complexity of fully automated agents in future models will necessitate sophisticated safeguards. The implementation of instruction hierarchy is seen as a step towards ensuring better security.

Continuous development and innovation in AI safety remain one of the biggest challenges facing the industry. However, OpenAI seems determined to stay ahead of the curve in this regard.