OpenAI Introduces Lockdown Mode to Limit Prompt Injection Risks

OpenAI has introduced a new optional security feature called Lockdown Mode, according to OpenAI's help documentation. The feature is designed to reduce the risk of data exfiltration from prompt injection attacks by limiting how ChatGPT interacts with the web and external services.

Lockdown Mode disables several capabilities, including live web browsing, deep research, agent mode, and file downloads. When enabled, web access is restricted to cached content, and ChatGPT cannot retrieve images from the web, though users can still upload images or generate new ones. The feature also restricts certain app and connector actions that could pose data exfiltration risks.

The mode is available to eligible personal accounts across Free, Go, Plus, and Pro tiers, as well as self-serve ChatGPT Business accounts. Workspace administrators can also enable Lockdown Mode for managed workspaces using role-based access controls. Lockdown Mode can be turned on through the Security settings under Advanced security.

OpenAI notes that Lockdown Mode does not eliminate prompt injection risks entirely, as malicious instructions could still appear in cached web content or uploaded files. However, it aims to significantly limit potential data exposure by blocking outbound network requests that could transfer sensitive information.