OpenAI's Lockdown Mode blocks prompt injection attacks for sensitive data

OpenAI has introduced Lockdown Mode, a new security feature designed to protect sensitive data from prompt injection attacks. Prompt injection occurs when malicious instructions are hidden in webpages or other content sources, tricking AI models into revealing confidential information or performing unintended actions. Lockdown Mode addresses this by disabling several high-risk capabilities: live web browsing (switching to cached content only), retrieval and display of images from the web (though image generation remains active), deep research, and agent mode. These restrictions significantly reduce the attack surface by preventing the AI from fetching live content that could contain hidden exploits.

However, OpenAI acknowledges that Lockdown Mode is not a complete solution. Prompt injections could still appear in cached web content or uploaded files, potentially affecting response behavior or accuracy. The feature is aimed at organizations and individuals handling sensitive data who prioritize stricter protection against data exfiltration. It is currently rolling out to self-serve ChatGPT Business accounts and eligible personal accounts, with broader availability expected. OpenAI explicitly states that Lockdown Mode is not intended for everyone—it trades off functionality for security, making it ideal for compliance-heavy workflows rather than general use.