OpenAI Strengthens ChatGPT with a Lock Mode Against Attacks

OpenAI recently announced the launch of a new feature aimed at enhancing the security of ChatGPT against prompt injection attacks. These attacks involve hiding malicious instructions within web pages and other content sources, thereby compromising the functionality of chatbots.

The Lock Mode, as OpenAI has named it, disables several features to reduce risks. Among these measures, live web browsing is disabled, limiting access to only cached content. Additionally, retrieving and displaying images from the web is also disabled, although image generation remains possible. Deep search features and agent mode are also turned off.

Despite these precautions, OpenAI acknowledges that ChatGPT may still be vulnerable to prompt injections, particularly through cached web content or uploaded files. These elements could still influence the behavior or accuracy of the responses provided by the chatbot.

The primary goal of Lock Mode is to decrease the likelihood that sensitive data is inadvertently shared. OpenAI specifies that this feature is not intended for all users. It specifically targets individuals and organizations handling sensitive data, who require enhanced protection against data exfiltration risks related to prompt injections.

Currently, OpenAI is rolling out Lock Mode for ChatGPT Business self-service accounts, as well as for certain eligible personal accounts.