New Delhi: OpenAI has introduced a new security feature for ChatGPT aimed at reducing the risk of sensitive information being exposed through prompt injection
attacks. The optional function is intended for those who want to restrict some internet-connected features and external integrations that may be used by malicious actors to illicitly access data from users and organisations and is known as ‘Lockdown Mode’.
The feature is being rolled out to eligible personal ChatGPT accounts (free, Go, Plus, Pro and self-serve ChatGPT Business accounts). OpenAI says Lockdown Mode is about stopping data exfiltrations, the unauthorised transfer of data, which is the last and most harmful step in a prompt injection attack.
What is Lockdown Mode?
Prompt injection attacks can be made by hiding instructions in web pages, documents, files or other content that is fed into the AI system. The instructions can be used to control the behaviour of a chatbot and, in certain instances, prompt it to disclose private data.
OpenAI stated that prompt injection won’t be prevented from showing up in text in Lockdown Mode. Rather, it’s designed to make the data more difficult to move off the platform if it happens to be successful. Taps into current capabilities such as sandboxing, monitoring systems, URL-based security and enterprise controls, the company said.
Features restricted when lockdown mode is enabled
Lockdown Mode disables some of ChatGPT’s features that require internet connectivity or external services.
The content of live web browsing is limited to cached content, so search results might be partial, missing, or stale. Deep Research and Agent Mode are completely turned off. Users will no longer be able to have ChatGPT download files for data analysis, but files that are uploaded by users will still be supported.
OpenAI also said that the retrieval of images from the web might be restricted, though image uploads and image generation will be possible when they are supported. Canvas-generated code cannot access the Internet as well.
Impact on apps, connectors and integrations
The new security setting alters the functionality of apps and connectors.
Connectors based on synced data will still work for personal use and business users of ChatGPT who are self-serving. But live connector access and writing are denied. If you have Lockdown Mode on, you can’t use features like shopping-agent experiences or financial integrations.
Managed enterprise workspaces will still maintain administrator controls, role-based access permissions and app-specific settings, OpenAI said. Administrators are advised to have a close look at the apps and actions that continue to be available for users in Lockdown Mode.
What remains unaffected?
Several key features of ChatGPT are unchanged despite the limitations.
OpenAI says Lockdown Mode won’t impact memory, conversation sharing, file uploading, image generation, or Lockdown Mode settings to use conversations to improve AI models. Access to the network within Codex also does not change.
The company also announced that Lockdown Mode is incompatible with Developer Mode. If one feature is enabled, the other will be disabled.
How to enable lockdown mode
The feature is available for users who have a personal account or a self-serve ChatGPT Business subscription. It’s available to users who have an eligible personal account or self-serve ChatGPT Business subscription.
- Open ChatGPT and go to Settings
- Select Security
- Go to Advanced Security
- Turn on Lockdown Mode
- Verify action when asked
After turning it on, there will be a status message above the chat composer. Lockdown Mode may also be turned off for individual chats without turning off Lockdown Mode.
Availability
Lockdown Mode will be accessible to all ChatGPT account types and workspaces, though users will need to be signed in to access the feature, according to OpenAI. Eligible personal accounts and self-serve ChatGPT Business customers are currently being rolled out.
