OpenAI has announced a new feature called Lockdown Mode for ChatGPT, specifically designed to counter cue injection attacks. This feature is aimed at users and organizations that need to handle sensitive data, with the goal of reducing the risk of data breaches.

Disable some network capabilities

When enabled, ChatGPT will not support real-time web browsing; only cached content will be accessible. Image retrieval and display will also be disabled, but image generation will remain available.

In-depth research and agency model are limited

OpenAI will also disable deep research and agent mode in this mode. The company stated that this reduces interference with model behavior when malicious commands are hidden in web pages, files, and other content.

Risks cannot be completely eliminated.

OpenAI also cautions that even with Lockdown Mode enabled, ChatGPT can still be vulnerable to tooltip injection. For example, cached webpage content or uploaded files may still contain malicious commands, thus affecting the accuracy of responses.

Additional information:This feature is currently being rolled out gradually to ChatGPT Business self-service accounts and eligible personal accounts.