aimode.news
Published on

OpenAIUnveils lockdown mode to protect user sensitive information from prompt injection

Authors

OpenAIUnveils lockdown mode to protect user sensitive information from prompt injection

OpenAI announced Lockdown Mode to protect user sensitive information from prompt injection。.

OpenAI Help Center

https://help.openai.com/en/clecles/20001061-lockdown-mode

Copyright © TechCrunch. All Rights Reserved.

https://techcrunch.com/2026/06/06/openai-unveils-lockdown-mode-to-protect-sensitive-data-from-prompt-injection-attacks/

Lockdown mode is an advanced security setting that is provided as an option to limit tools and features that can be connected to the web or external services of OpenAI products. If you briefly explain the lockdown mode, it means “restricting requests to external networks to reduce the risk of data leakage caused by prompt injection。.” However, please note that some useful features will be disabled or。 as a fee to improve security.

For this reason, OpenAI is not suitable for all users. “It’s designed for individuals and organizations that need more stringent protection from data leakage risks。 with sensitive data and prompt injection .”

Lockdown modes are available in all account types and workspaces, but they are scheduled to be deployed se。tially, so some users may not see the lockdown mode in the settings screen.

Lockdown modeChatGPTFrom the setting to the "Advanced Security" menu. Lockdown mode and developer mode cannot be enabled at the same time. If the lockdown mode is enabled, the status message will appear on the composer. To turn off only a specific chat, select Manage from the status message and select OFF with this chat. You can also open "Other Options" and select "Lock Down" and select "Disable".

In the first place, a prompt injection attack is a type of cyberattack against AI that操作e AI to leak sensitive information and spread false information by disguising malicious input as a legitimate prompt.

OpenAI continuously enhances multi-layer security and safety systems to protect your sensitive information from prompt injection。, and lockdown modes are based on protection across the entire AI model, product and system level. This website includes enterprise control such as "sandbox", "protection against URL-based data leakage", "survey and force", "roll-based access and audit log".

Lockdown mode is designed to prevent the final stage of data leakage due to prompt injection。 by limiting transmission network requests that may transfer sensitive data to attackers. ChatGPTcannot prevent prompt injection to the content that is processed. For example, a prompt injection attack may occur in cached web content or uploaded files, and it is important to note that it may affect response behavior and accuracy.

If the lockdown mode is enabled, the following functions of the OpenAI product will be disabled.

Live web browsing: Web browsing is limited only to access to cached content. If the search result is limited, it may be unavailable or old information.

Image support: ChatGPT may not be able to display images in normal responses or retrieve images from the web. Users can still upload image files. However, image generation function is still available if other functions are available.

Deep Research:。 Research

Agent mode: Agent mode is disabled.

Canvas Network Functions: Users cannot approve access to the network by the code generated by the canvas.

File download: ChatGPT will no longer be able to download files for data analysis. However, the file uploaded manually can be processed by ChatGPT.

In lockdown mode, functions such as memory usage, file upload, conversation sharing, and conversation content are used to improve the AI model are not changed. Many of these settings can be configured individually by the workspace administrator.

In addition, if you enable lockdown mode according to OpenAI, the setting for whether to use the conversation contents for training AI models will not change.

・Re ・ articles

ChatGPT’s memory is smartly updated to improve the ability to adapt to users’ preferences and ability to forget old information properly - GIGAZINE

ChatGPT 5.5 Pro runs a doctoral level mathematics study in one hour, noting that mathematics will change the minimum line of human studies - GIGAZINE

The Government of to Provide ChatGPT Plus to Free Nationals for One Year of AI Course Completed - GIGAZINE

AI agent "workspace agents" for ChatGPT is now available - GIGAZINE

ChatGPT is fast learning, but it can be hard to rest in long-term memory - GIGAZINE

GIGAZINE

OpenAIUnveils lockdown mode to protect user sensitive information from prompt injection | aimode.news