OpenAI introduces Lockdown Mode to shield users from prompt injection attacks

OpenAI is rolling out a new optional security feature called Lockdown Mode, designed to protect users from prompt injection attacks. Igor Bonifacic reports for Engadget that the feature is aimed at people and organisations handling sensitive data.

Prompt injection is a form of social engineering targeting AI chatbots. Attackers hide malicious instructions on webpages or in other content. When an AI system reads that content, it may follow those hidden commands without the user’s knowledge.

What Lockdown Mode does and does not do

OpenAI describes Lockdown Mode as a last line of defence. It does not prevent malicious instructions from appearing in content that ChatGPT processes. Instead, it limits the network requests ChatGPT can make, reducing the risk that an attacker could extract sensitive data from an account.

Enabling the feature restricts several ChatGPT functions:

• ChatGPT cannot retrieve images from the internet or display images in responses
• The chatbot cannot download files to analyse, though users can still upload documents manually
• Deep Research and Agent Mode are disabled entirely

Memory, file uploads, conversation sharing, and model training settings remain unaffected by Lockdown Mode, according to OpenAI.

The feature is available to all users, including those on ChatGPT’s free tier. To enable it, users open ChatGPT’s settings, select “Safety and security,” then find “Lockdown mode” under “Advanced security.” It can also be turned off temporarily for individual conversations.

Alongside Lockdown Mode, OpenAI is introducing an active session manager. It shows users which devices and browsers are currently logged into their account and allows them to sign out of individual or all sessions at once. OpenAI notes that a full sign-out can take up to 30 minutes to complete.